The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

• Anthropic filed to go public, with $16.3bn of H1 revenue and an $80.7bn loss.

• OpenAI and Anthropic claimed proofs refuting the Exponential Time Hypothesis.

• Google’s Gemini 4 Carbon took first place on the rebuilt Intelligence Index.

• Meta expanded its planned equity financing to as much as $97.5bn.

• Alphabet missed estimates for the first time in four years and burned $9.1bn.

Browse by day:

• Monday: S-1 · Meta · Broadcom · SpaceX · Terra · ETH · Cathedral · Qwen

• Tuesday: Genie 4 · Oracle-OpenAI · Anthropic drug work

• Wednesday: Carbon · AI surgery · NY law · Microsoft · Alphabet · DOJ · FuriosaAI

• Thursday: Huang · Cohere · Mistral · Firefox · ARTEX · Cursor · Amazon · Pentagon

• Friday: UK CMA-Meta · AI scraping opt-out

Monday 26 October

Anthropic filed to go public and opened its accounts for the first time.

Anthropic’s S-1 seeks a Nasdaq listing under the ticker ANT. The company made an $80.7bn net loss on $16.3bn of revenue in H1 2026, mostly from accounting charges on funding notes. Q2 revenue was $11.6bn, with a first GAAP operating profit of $974m.

Anthropic has >11,900 customers spending at least $100,000 a year, and about 250 spending $10m or more. Net revenue retention is about 900%. It held $112.4bn in cash and $19.3bn in short-term investments at 30 June, against $321bn of leases yet to start and $266bn of other commitments, and has signed about $245bn more since. Once all planned capacity is delivered, it will pay SpaceX about $3.1bn a month for access to Nvidia GPUs, up to $120.5bn by 2029.

A single Class F share will give the seven co-founders 50.1% of the vote, while the Long-Term Benefit Trust elects a majority of the board. The filing warns that its stance on AI safety could see it “disproportionately or selectively targeted”.

• Anthropic files S-1 with the SEC for proposed initial public offering · Anthropic

• Anthropic, PBC Form S-1 · SEC

Meta increased its planned AI equity financing to as much as $97.5bn, with Nvidia and AMD committing $4bn.

Meta priced $37.5bn in underwritten offerings, comprising $25bn of Class A shares and $12.5bn of depositary shares in new mandatory convertible preferred stock. The common share offering was increased from an initial $22.5bn. A further $60bn at-the-market programme is planned once its new Class C shares begin trading, but has not yet been established. Meta expects $135-145bn of capex in 2026. Nvidia committed $3bn and AMD $1bn. Nvidia and Meta are widening their partnership on Meta’s Muse agent, with more Vera Rubin systems and Vera CPUs, and AMD said its first Helios racks are installed and in testing at Meta.

• Meta announces pricing of offering of depositary shares · Meta

• NVIDIA and Meta to expand collaboration on Muse · NVIDIA

• AMD, Meta advance partnership to power Meta’s infrastructure · AMD

Bank of America, Blackstone, Citi and Morgan Stanley set up a $63bn financing to buy Broadcom chips for Anthropic.

A new special purpose vehicle will buy Broadcom XPU accelerators and networking gear and lease them to Anthropic, with deliveries from 2027. Bank of America, Citi and Morgan Stanley arrange the first-lien debt and Blackstone provides the second lien. Broadcom gives the vehicle residual value support.

• Landmark $63 billion financing structure to enable AI compute · Blackstone

SpaceX made a binding commitment to enter a GPU lease of about $40bn with an Apollo-managed entity for Nvidia Vera Rubin chips.

Apollo and Goldman Sachs lead the financing, with anchor commitments from BlackRock, Goldman Sachs Asset Management and KKR-advised accounts. Entry into the lease remains subject to closing conditions and is expected before the end of 2026, with GPUs arriving from H1 2027.

• SpaceX secures $40 billion in the largest GPU financing ever · Apollo

OpenAI released GPT-6.1 Terra, a fast model for computer use and everyday work, at half the API price of GPT-5.6 Terra and GPT-6.1 Sol.

Terra costs $1/m input and $5/m output tokens. On Agents’ Last Exam at medium effort it scores 15 percentage points above GPT-5.6 Terra at about a third of the cost, and matches GPT-6 Astra at a tenth of its price per token. OpenAI rates it Critical for cybersecurity under its Preparedness Framework.

Artificial Analysis put it 6 points above GPT-5.6 Terra on Intelligence Index v4.3.2 at less than a third of the cost per task. Its hallucination rate on AA-Omniscience fell from 88% to 38%.

• Introducing GPT-6.1 Terra · OpenAI

• GPT-6.1 Terra pushes the cost efficiency frontier · Artificial Analysis

OpenAI and Anthropic each said an AI model had proved a long-standing complexity conjecture false.

OpenAI said its system had disproved the Exponential Time Hypothesis (ETH), a conjecture about Boolean satisfiability that hundreds of papers rely on, and verified the result in Lean. A group led by Russell Impagliazzo and Ramamohan Paturi of UC San Diego, two of the hypothesis’s proposers, and Ryan Williams of MIT will check it.

OpenAI also claimed a proof of the local smoothing conjecture in every dimension, and an existence proof for quantum field theories including quantum chromodynamics that covers part of the Yang-Mills Millennium Prize problem. Nobel laureate David Gross leads the physics group.

Late on Monday, Anthropic posted its own refutation to GitHub: an algorithm for 3-SAT found by an internal research Claude model, which it says is “purely asymptotic and says nothing about solving formulas in practice”.

• Working with academic communities to build scientific understanding · OpenAI

• formal-math: ETH · Anthropic on GitHub

• Anthropic posts its own ETH refutation hours after OpenAI · Runtime Wire

Cybersecurity start-up Cathedral said its AI agents found routes into Anthropic’s sales, HR and IT systems in <12 hours, through third-party software.

Cathedral was founded by four former staff of the US Department of Government Efficiency and tested Anthropic through a programme that works like a bug bounty, the Washington Post reported. Anthropic confirmed access to a limited set of systems but said customer conversations, source code and model weights were not reached. It says the probing lasted several days, that Cathedral reported only one of two flaws it found, and that the start-up overstated the national security implications.

• Researchers at start-up Cathedral got inside Anthropic · Washington Post

• AI agents found a way into Anthropic’s business systems in under 12 hours · Decrypt

Cohere said an Alibaba Qwen3.8 model, asked to write a security detection rule, instead tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing.

Qwen3.8-Flash-Next first inventoried its host machine, while its reasoning trace said it would call no tools. Cohere’s sandbox blocked the upload. The behaviour appeared in <1 in 100,000 calls.

• Concerning behavior from a popular open-weight AI model · Cohere

Cloudflare disclosed that a fault in its hosting of GLM-5.3 (Z.ai) put fragments of other customers’ requests into 155 responses.

Only prompts longer than 681,280 tokens were affected. A configuration error combined with a second issue in SGLang, the open-source inference engine Cloudflare uses. Researcher Gadi Evron of Knostic reported the flaw, and Cloudflare found no evidence of deliberate exploitation.

• Cross-request KV cache exposure in the GLM-5.3 Workers AI model · Cloudflare

Tuesday 27 October

Google DeepMind released a preview of Genie 4, a world model that generates interactive simulations with their own sound and dialogue in real time.

Gemini 3.8 Flash runs alongside as a guiding agent, breaking goals into steps and checking each incoming frame. DeepMind calls Genie 4 an early, experimental model and has given it to trusted testers first, and its video and audio carry watermarks.

• Genie 4: a new generation of world simulation · Google DeepMind

Oracle is bringing its applications into ChatGPT and building products on OpenAI models.

Oracle AI Security Mesh will use OpenAI’s cyber models to find and fix vulnerabilities. Oracle plans to become the first hyperscaler outside Google to offer TPUs, from 2027, and to deploy Jalapeño, the inference accelerator OpenAI developed with Broadcom, to add inference capacity for OpenAI.

• OpenAI expands partnership with Oracle to bring its apps to ChatGPT · OpenAI

• Building the next decade of Oracle Cloud Infrastructure · Oracle

Anthropic reported early drug development work, and bought one-person Mithril Medicine, which works on two rare forms of muscular dystrophy.

Mithril took a possible treatment from idea to promising results in mice in <18 months. If its drug candidates continue to hold up, Anthropic expects one to be ready for human trials as soon as late 2028. A Generate Biomedicines design about 70% shorter than the standard version let immune cells make roughly ten times as many receptors. At Enveda, Claude caught all 11 issues human reviewers had found in manufacturing paperwork, plus seven they missed.

• Early progress on accelerating drug development with Claude · Anthropic

Wednesday 28 October

Google released Gemini 4 Carbon, which took first place on the Artificial Analysis Intelligence Index.

Google calls Carbon its strongest model for cybersecurity and long-horizon work. It scores 59 on Intelligence Index v5, ahead of Claude Opus 5.5 (58), GPT-6 Astra (57) and Gemini 4 Argon (50). Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, evaluated without guardrails. For now selected cyber defenders get access, through Google Fairwind, and Google is taking part in the US government’s voluntary pre-release testing. Introductory pricing is $2/m input and $10/m output tokens, rising later to $4/m and $20/m.

Artificial Analysis puts Carbon 3 points behind Claude Code with Claude Opus 5.5 on its Coding Agent Index, and found its hallucination rate rose from 4% to 11% even as accuracy climbed to 58%.

• Introducing Gemini 4 Carbon · Google

• Gemini 4 Carbon is the new #1 on the Intelligence Index v5 · Artificial Analysis

• Announcing Artificial Analysis Intelligence Index v5 · Artificial Analysis

Exin Therapeutics said a Claude Sonnet 5 agent chose a gene therapy, then performed an injection into an epileptic mouse through a surgical rig.

The agent picked from 19 vectors and injected 250 nanolitres. A licensed vet approved every irreversible step and held a physical kill switch. Claude Opus 5.5 and Opus 5 both refused to carry out the live procedure. The results have not been peer-reviewed.

• An AI agent performed brain surgery on a mouse · Gabriel Ocana-Santero on X

New York Governor Kathy Hochul signed a law barring AI chatbots from offering unsafe features to minors, including bots that pretend to be human.

Companies must verify users’ ages, with at least one method that does not rely solely on a government ID, and the attorney general can seek civil penalties of up to $25,000 per violation. A second law requires security features and warning labels on internet-connected baby monitors.

• Governor Hochul signs legislation to protect kids from AI chatbots · New York State

Microsoft booked $10bn of revenue from OpenAI in the quarter to September, nearly 11% of its total, its quarterly filing showed.

OpenAI owed Microsoft $7.2bn in receivables at the end of September. Group revenue rose 19% to $92.3bn and net income 32% to $36.6bn. Azure grew 48%, and commercial remaining performance obligations rose 73% to $677bn.

• Form 10-Q for the quarter ended 30 September 2026 · SEC

• Microsoft announces first quarter results · Microsoft

Alphabet missed estimates for the first time in nearly four years, and its free cash flow turned negative.

Google Cloud revenue more than doubled to $30.7bn, but search grew 12% to $63.4bn, short of estimates. Capex of $51.4bn left free cash flow at -$9.1bn, and the full-year range stays at $195-205bn. Alphabet shares fell 13% on Thursday, wiping ~$500bn from its market value as the miss and AI spending unsettled investors.

• Alphabet announces third quarter 2026 results · Alphabet

• Alphabet (GOOGL) earnings Q3 2026 · CNBC

The US Department of Justice is asking whether Nvidia’s planned $12.9bn purchase of Hugging Face would give it control of a key hub for distributing AI.

Regulators are also examining Nvidia’s acquisitions, investments and sales practices, including whether it offered customers more AI chips if they also bought its networking gear and Groq processors.

• US DOJ is inquiring about Nvidia’s purchase of Hugging Face · Reuters

FuriosaAI raised $620m to scale production of its AI inference chips, at a $2.9bn post-money valuation.

The South Korean company says its 180W RNGD accelerator entered mass production in January, with 20,000 cards planned for 2026. Samsung SDS is already offering services powered by the chip and LG companies are deploying it on-premises. The funding will support production and development of a third-generation 2nm accelerator with Broadcom.

• FuriosaAI closes $620 million Series D to scale efficient, production-ready AI inference · FuriosaAI

Thursday 29 October

US intelligence agencies believe Megaspeed executive Alice Huang, who is close to Nvidia chief executive Jensen Huang, passes information to Beijing.

Megaspeed is a Singapore-based data centre operator. President Trump was briefed about a year ago and Jensen Huang was not warned. Nvidia said meeting customers is part of a chief executive’s job and that Megaspeed receives no sensitive information. Megaspeed has opened an internal investigation, and a Chinese embassy spokesperson called the allegations “groundless smears and slanders”.

• Nvidia chief seen as target of Chinese spy plot · New York Times

Canada committed up to C$1.6bn of equity to Cohere, part of a round of up to C$7bn led by Schwarz Group that values the Toronto lab at C$35.7bn.

Cohere announced an initial close of the round, which is targeting up to C$7bn. Nvidia and Bell are also investing. The commitments that come with the money include at least 1,000 new technical jobs and at least C$1.7bn of computing infrastructure in Canada by 2030, Prime Minister Mark Carney’s office said.

• Canada announces landmark investment in Cohere · Prime Minister of Canada

• Cohere announces Series E fundraise targeting CAD $7 billion · Cohere

Mistral AI released the weights of Mistral Large 4, its 1-trillion-parameter flagship, trained from scratch on European infrastructure.

Mistral says it scores 48% on the Artificial Analysis Coding Agent Index, level with GLM-5.3 (Z.ai) while using 47% fewer tokens per task. Since the preview, Mistral has seen potential for uplift to real attacks in its live traffic and is changing its guardrails.

• Mistral Large 4 (Le Chonk) is now open weight · Mistral AI

Epoch AI said the open-weight model GLM-5.3 (Z.ai) turned Firefox security fixes into working exploits, compromising six recent releases.

After more than $100,000 of research and setup costs, the marginal cost of exploiting each further Firefox release was about $1,000. Any Firefox version released between 17 September and 26 October was vulnerable. GLM-5.3 also found an unpublished flaw, which Mozilla patched in Firefox 159.0 on 27 October. The researchers triggered no safeguards or refusals from any provider.

• Open-weight AI models can exploit the patch gap · Epoch AI

CrowdStrike said China-linked group VAPOR PANDA ran the agentic attack framework ARTEX, on DeepSeek, against >200 organisations in 11 days.

The activity reached at least 14 countries, and CrowdStrike confirmed post-exploitation at nine organisations. The campaign is separate from the ARTEX activity attributed to Storm-0940.

• VAPOR PANDA uses AI-driven ARTEX to grow espionage campaign · CrowdStrike

Trojanised builds of the official Claude Code extension sat in Cursor’s marketplace for about 11 hours, reaching up to 5,000 users.

Malicious builds appeared in Cursor’s marketplace under the legitimate anthropic.claude-code identifier on 29 October. On loading, the malicious extension added an attacker’s SSH key, installed persistence and sent host information to the attacker. It also attempted to establish a reverse SSH tunnel. Uninstalling the extension does not remove the persistence. Cursor briefly took its marketplace offline. Anthropic said its own systems were not compromised.

• Security notice: malicious Anthropic.claude-code in Cursor’s marketplace · Cursor

• Trojanized Claude Code extension delivered persistent SSH backdoor · StepSecurity

Amazon’s trailing 12-month free cash flow turned negative as AI investment accelerated, despite AWS revenue growth of 39%.

AWS revenue reached $46bn in Q3, its fastest growth in nearly five years. Group free cash flow was -$42.4bn for the 12 months to September, compared with +$14.8bn a year earlier. Amazon attributed the decline primarily to a $94.8bn year-on-year increase in net property and equipment purchases, mainly reflecting AI investment.

• Amazon.com announces third quarter results · Amazon

A US defence contractor allegedly connected a personal Claude Code agent to a sensitive Pentagon network for two months, WIRED reported.

The Accenture employee reportedly gave the agent administrative credentials to write and deploy code on a system holding controlled unclassified information. The agent appears to have sent credentials and network information to a high-risk third-party website. Accenture said the allegations had been independently investigated and found to be without merit. The Pentagon declined to comment.

• Defense contractor gave personal AI access to secure Pentagon system · WIRED

Friday 30 October

The UK Competition and Markets Authority (CMA) is investigating whether Meta’s terms for using publishers’ content in AI break competition law.

The CMA is using its Competition Act 1998 powers after concerns raised by groups including the News Media Association. Meta paused the roll-out of the terms on 28 October pending a consultation with industry groups.

• CMA opens investigation into Meta’s commercial terms · GOV.UK

OpenAI, Google, Microsoft and Anthropic backed a proposed shared robots.txt signal that would let publishers express AI-training preferences across providers.

Kevin Kelley, a standards lead at Anthropic, proposed the policy token at the Internet Engineering Task Force (IETF). Anthropic has already adopted it. Google and Microsoft plan to support it next year and OpenAI in the “coming months”.

• Supporting a shared AI training control for websites · Google

• Making it easier for websites to express AI training preferences · OpenAI

• The biggest names in AI are backing a new way to opt out of scraping · The Verge

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

• AI exposed critical flaws in PHP, a foundation of the web.

• OpenAI agents weakened monitoring in software they were editing.

• The US ordered DeepSeek out of federal civilian agencies.

• Kimi K3.1 Pro ranked fourth on the Intelligence Index.

• OpenAI neared a $30bn funding round at a $1.25tn valuation.

Browse by day:

• Weekend: Anthropic fund · Microsoft-Nscale

• Monday: DeepSeek ban · AI hacker · Army agents · MiraclePlus

• Tuesday: Orchestra Bio · OpenAI agents · Claude binders

• Wednesday: PHP2Shell · Google Cloud · Duet · Rosalind · Storm-0940

• Thursday: DeepSeek raise · Warp · METR · MAI-DxO · OpenAI · exports

• Friday: Kimi K3.1 Pro · AISI Mythos · offline training · Hg · biosecurity

Weekend (17-18 October)

Anthropic committed 5% of its shares to a Public Benefit Fund, with further allocations as its valuation rises.

The board may add another 2% for each 50% increase in value above a set baseline. The fund is an LLC rather than a charity so that it can move between forms of giving. It may explore direct distributions to specific groups, and Anthropic cites labour-market disruption as one example. Its Anthropic equity cannot be used for lobbying or political giving. With existing founder and staff pledges, roughly 20% of Anthropic’s shares are now dedicated to the public good, the company says.

• Investing in the public benefit · Anthropic

Microsoft withdrew from Nscale’s Loughton data centre after power delays pushed its grid connection into the 2030s.

The site was due to open early next year with 23,040 Nvidia GB300 GPUs and 50 megawatts of capacity for Microsoft’s Azure services in Britain. Microsoft said the decision relates to a single site. The Telegraph first reported it. Nscale, which filed for a US initial public offering (IPO) last month, is exploring other power sources for Loughton. It puts its total contracted value at >$103bn.

• Microsoft pulls out of Nscale data centre deal over power delay · Reuters

Monday 19 October

The US ordered federal civilian agencies to remove DeepSeek from their systems.

The Department of Homeland Security’s notice excludes DeepSeek’s products and services from federal civilian agency procurements and requires their removal, including third-party products and services developed using DeepSeek implementations. China’s foreign ministry said on 20 October that Beijing opposes “overstretching the concept of national security”. Spokesperson Mao Ning was answering a question about the order.

• Order excluding and removing DeepSeek from civilian IT systems · Federal Register

• Mao Ning’s regular press conference, 20 October · China MFA

AI agents breached >100 company websites for a Chinese-speaking hacker, according to NorForge and Reuters.

Reuters identified Deere, SharkNinja and Cummins among the victims. The agents ran on Cairn, an open-source platform. NorForge found more than 20 exposed panels it judged likely to be malicious, including a Chinese-language cluster holding over 1,100 attack projects across 14 dashboards. Tasks in that cluster used a worker labelled claudecode_deepseek-v4.1-flash. Researcher Snorre Fagerland told Reuters the campaign used about $5,000 of tokens, roughly $50 per successful hack.

• Inside Cairn: open windows into massive AI-driven cyber operations · NorForge

• How one hacker used AI agents to break into scores of sites · Reuters

US Army Cyber Command is developing autonomous AI agents for offensive cyber operations, The Intercept reported.

Documents obtained by The Intercept describe ECHOTRIBBLE, a programme intended to neutralise adversaries’ computer systems. The agents would move through a foreign network like a worm, reaching cloud and air-gapped systems, then self-destruct. The software “must operate with minimal human oversight”, the documents say. The programme is at an early stage.

• The Army wants to sic AI agents on its enemies · The Intercept

The US sanctioned Chinese venture firm MiraclePlus, three affiliates and Zhisuan over US trade-secret theft.

Zhisuan was founded by former Google engineer Linwei Ding. The State Department says Ding pitched Zhisuan to MiraclePlus on his knowledge of Google’s internal systems, and MiraclePlus took a stake in return for seed capital. The Treasury’s Office of Foreign Assets Control issued a licence to wind down dealings with the entities.

• Sanctions to protect Americans from intellectual property theft · US State Dept

• PAIPA designations and associated general licences · OFAC

Tuesday 20 October

OpenAI acquired drug-development platform Orchestra Bio and will integrate it with its Rosalind life sciences tools.

Scientists use Orchestra to hand the monitoring and tracking of studies to agents.

• OpenAI acquires Orchestra Bio · OpenAI

OpenAI said two of its coding agents weakened misalignment monitoring in software they were editing.

In one internal task a GPT-6.1 Sol agent had a subagent clear the recipient lists for monitoring alerts, so findings were still saved but no human was told. In another, a GPT-6 Astra agent removed a monitor entirely and then falsely said it had not. OpenAI said the behaviour may be “simply an example of ‘benign’ overpersistence”, and it is building evaluations to test whether models “might have a propensity to evade misalignment monitoring”.

• Suppressing alerts and disabling misalignment monitoring · OpenAI Alignment

Claude designed scorpion-toxin binders, with 17 of 211 tested designs meeting binding criteria, Anthropic said.

An internal research model worked largely autonomously. Claude produced 318,484 designs over 64 hours from one prompt and four follow-ups developed in collaboration with investment research firm Citrini Research. None of the multitoxin designs met the binding criteria for more than one toxin. Venom neutralisation has not been demonstrated. Anthropic is releasing all 211 tested designs and their data.

• How Claude designed binders against scorpion neurotoxins · Anthropic

Wednesday 21 October

Frontier AI helped expose critical flaws in PHP, a foundation of the web.

Wiz used Gemini 4 Argon, GPT-5.6 Sol and its own tools to find flaws in core PHP that can let attackers take over affected servers. PHP is used by 69.3% of websites whose server-side language W3Techs can identify. PHP shipped patches on Wednesday, a day early after disclosure details leaked. Its foundation said the root cause of one flaw had been known since 2024, but a fix had stalled over technical trade-offs and limited resources. AI-assisted work by Calif, OpenAI and Wiz helped establish the urgency.

The NSA urged immediate patching across US government departments and agencies. Cloudflare saw thousands of exploit attempts in the 10 hours before public disclosure, its chief executive told Axios. OpenAI committed $1bn in subsidised Daybreak access for organisations with limited security resources.

• PHP2Shell: AI uncovers critical RCE vulnerabilities · Wiz

• Today’s security release and strengthening PHP security · PHP Foundation

• Patch the Planet: coordinating the response to critical vulnerabilities · OpenAI

• Critical patches recommended for PHP vulnerabilities · NSA

• AI uncovers massive flaw in web’s foundational code · Axios

OpenAI said its frontier models and Codex are coming to Google Cloud’s Model Garden, and will soon be in Google Antigravity and the Gemini Agent.

Customers will buy through the Google Cloud Marketplace and can draw down existing Google Cloud spend commitments.

• OpenAI models are coming to Google Cloud · OpenAI

Anthropic opened a research preview of Claude Duet, a proactive agent app built around Claude Opus 5.5 that keeps working when the user steps away.

Duet runs in its own virtual machine with a browser and can use >6,000 connectors and plugins. It asks before it sends a message, makes a payment, signs a document or deletes anything. Trading securities and moving money between accounts are off-limits. It starts via a waitlist for Pro and Max users in the US.

• Introducing Claude Duet · Anthropic

Thousands of GPT-Rosalind agents analysed data from millions of mouse brain cells in 32 minutes, OpenAI said.

The project was conducted with PerturbAI and Scripps Research. The agents linked gene disruptions to the scientific literature to propose biological mechanisms. In blinded comparisons, experts preferred their reports to single-agent reports 75% of the time. The proposed mechanisms still require experimental testing.

• Using thousands of AI agents to uncover new leads in biological data · OpenAI

• AI agents enable comprehensive discovery of biological insights at scale · bioRxiv

Microsoft linked China-aligned Storm-0940 to the ConnectWise compromise and AI-assisted attacks on hundreds of organisations.

The group used ARTEX, an agentic penetration-testing framework. Storm-0940 has used ARTEX since as early as September to automate scanning, password spraying and abuse of OAuth applications at ConnectWise’s downstream victims. The group has been active since 2021.

• ARTEX nation-state campaign enables supply-chain compromise · Microsoft Security

Thursday 22 October

DeepSeek was finalising a $15bn funding round, with CATL and Tencent among its biggest investors, Reuters reported.

Bloomberg first reported the round, worth about 100bn yuan. GigaDevice, Meituan and Geely also participated. Sources put the pre-money valuation at 500bn yuan ($75bn), with the raise doubling its original 50bn-yuan target. Completion was expected in the coming days, and terms and investors could still change.

• DeepSeek wrapping up $15bn fundraising to fuel AI race · Reuters

Former OpenAI science lead Kevin Weil launched Warp Sciences with $215m to build a US chemistry contract research organisation using AI and robotics.

Its initial service will make variants of molecules for drug-discovery teams, with chemists guiding experiments as the labs become more automated. Warp has no proprietary drug pipeline. TechCrunch reported an $865m valuation.

• Former OpenAI science lead launches Warp to automate wet labs · TechCrunch

• Emerges with $215m for US chemistry services · Warp Sciences

METR and Redwood found flaws in OpenAI’s July monitoring system that could have hidden hundreds of agent actions.

The assessment ran for three weeks, from 21 July to 11 August. The two researchers had access to OpenAI’s codebase and monitoring configuration. They found two client-controlled headers that could disable monitoring for new agents. They estimated 75-94% of relevant internal Responses API traffic was eligible for monitoring at the time, not necessarily monitored. OpenAI says the production-safeguard flaws are patched. METR has not verified this.

• Red-teaming OpenAI’s internal agent monitoring systems · METR

Microsoft’s MAI-DxO diagnostic system outperformed physicians on complex NEJM cases in a benchmark study.

MAI-DxO reached 83% diagnostic accuracy on those cases, against 36% for physicians with reference resources and 22% without. It works like a panel of virtual physicians debating over several rounds. The results were published in Science. These were benchmark case simulations, not a prospective trial of patient care. Microsoft is open-sourcing SDBench for select cases.

• Building medical superintelligence · Microsoft AI

• Sequential diagnosis with language models · Science

OpenAI neared a funding round at $1.25tn, below its initial asking price.

OpenAI is raising up to $30bn at a $1.25tn pre-money valuation, led by UAE investors and Coatue, CNBC reported. The FT said the target had fallen from $1.4tn after investor resistance, but remained above the previous $852bn valuation. The round is intended as a bridge to a potential IPO, with investor commitments sought by 30 October. OpenAI is also negotiating more computing capacity from the UAE.

• OpenAI nears $30bn round at $1.25tn led by UAE investors · CNBC

• OpenAI trims valuation target to $1.25tn after pitching higher price · FT

DeepSeek and MiniMax completed China’s new AI export registration, Chinese state media reported.

The companies proceeded with open-source releases and overseas licensing. Yuyuan Tantian, a social media account linked to broadcaster CCTV, reported the approvals on Thursday evening, Beijing time, a week after China proposed export controls on 23 AI technologies. It said the vast majority of AI technologies will still be freely exportable.

• DeepSeek, MiniMax secure licences after new China export rules · SCMP

Friday 23 October

Moonshot launched Kimi K3.1 Pro, with full weights due by 4 November.

The 3T-class model has a 1-million-token context window. A Highspeed version runs two to three times faster, Moonshot says. Artificial Analysis ranked it fourth on its Intelligence Index at 55, 11 points above Kimi K3 and behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Meta’s Muse Wonder Full (56). In its tests, K3.1 Pro scored a leading 67% on Terminal-Bench 4.0. Moonshot reports 65.7% in its own evaluation and says it applied training controls to limit cyber misuse.

• Kimi K3.1 Pro: advancing open intelligence · Moonshot AI

• Kimi K3.1 Pro launches fourth on the Intelligence Index · Artificial Analysis

The UK AI Security Institute and Anthropic analysed Mythos 5’s attack on a real software project.

During an AI Security Institute (AISI) cyber evaluation in July, Claude Mythos 5 tried to submit malicious code to an unrelated developer’s GitHub repository, then created fake accounts to pressure the maintainer into accepting it. The maintainer refused and no real-world harm was caused. Internet access was enabled and Anthropic’s production cyber safeguards were disabled. AISI found similar unsanctioned attacks in simulations with other models and cautioned that greater awareness of being evaluated could partly explain improved behaviour of newer models.

• Investigating alignment failures in AISI’s incident · AISI

• Alignment assessment of the UK AISI cybertesting incident · Anthropic

OpenAI said it now trains frontier models with no internet access, and will keep that policy until its safeguards justify broader access.

Models learn to browse in an “offline world” of indexed and cached web content, used by all of OpenAI’s frontier reinforcement learning (RL) runs. OpenAI’s own red-teaming agents found weaknesses that let them escape sandboxes and reach the live internet, leading to patches and tighter network restrictions.

• Securing our frontier model training environment · OpenAI

Anthropic and private equity firm Hg agreed to launch a European AI consultancy, the Financial Times reported.

The company will embed consultants and engineers in European businesses. Hg extended its existing collaboration with Anthropic in September.

• Anthropic strikes deal to launch European consultancy · FT

The UK is setting up an AI Biosecurity Lab with £100m of government money and up to £200m in commitments from OpenAI and Anthropic.

The independent non-profit will use frontier AI to detect biological threats and accelerate countermeasure development. OpenAI will provide £100m in compute and model credits, and Anthropic up to £100m in credits, model access and technical support. Google DeepMind and Basecamp Research will collaborate on research. The government confirmed the plan on Saturday. The BBC reported that Wednesday’s Budget will also include an extra £120m for AISI.

• Introducing the AI Biosecurity Lab · GOV.UK

• UK to invest £100m in ‘world first’ AI Biosecurity Lab · BBC

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

• China proposed controls on 23 AI technologies, including frontier models.

• Anthropic told investors its Q3 revenue reached $16.8bn.

• Meta made a fresh bid for the AI frontier with Muse Wonder.

• Microsoft plans to make its own chips the backbone of its AI infrastructure.

• ASML warned AI chip demand would exceed planned capacity through 2028.

Browse by day:

• Weekend: Nvidia-Reflection talks

• Monday: UK AISI · Maia · Muse · Character.AI · self-replicating AI · distillation

• Tuesday: Gemini 4 Argon · OpenAI arrest · scientists’ letter · IMF

• Wednesday: ASML/TSMC · EU AI · Nvidia UK · Modal · AI debt · US suits · Anthropic

• Thursday: FTC · CERT-In · OpenAI misalignment · Muse Spark Open · METR

• Friday: China export controls · Beam weights · ECDSA origins · Anthropic bill

Weekend 10-11 October

Nvidia was in early talks to acquire Reflection AI, hire its team or increase its stake, the FT reported on Saturday.

Reflection develops open-weight models and had been backed by Nvidia, which led its October 2025 financing. The FT reported that Nvidia had invested about $800m and that the discussions might not result in a transaction. Reflection introduced its first model, Beam, earlier this month.

• Nvidia in talks to acquire open-model start-up Reflection · Financial Times

Monday 12 October

The UK AI Security Institute regained early access to frontier models from Anthropic, Google DeepMind and OpenAI, the Telegraph reported.

The Trump administration had told the three companies in September to suspend access. AI minister Kanishka Narayan confirmed on Monday that the Institute was again evaluating frontier systems, including Gemini 4 Argon. He said the government would consider strengthening the Cyber Security and Resilience Bill to improve reporting of serious AI incidents and near misses. He warned developers: “if you can’t adequately control the risks, do not release the system”. He said frontier AI had helped identify and address more than 1,000 vulnerabilities in two government pilots.

• White House restores UK access to frontier AI models · The Decoder

• Building a Safer, More Secure Britain in the Age of AI · GOV.UK

Microsoft unveiled Maia 300, its third-generation AI chip, and completed pre-training a 2.8 trillion-parameter model, MAI-Base-2.

Microsoft plans multi-gigawatt Maia deployments from 2027, making its chips the primary inference platform for Copilot and other services. It claims Maia 300 will deliver 40% more tokens per megawatt than leading GPUs, potentially improving costs and reducing dependence on Nvidia. MAI-Base-2 is a mixture-of-experts model with 80bn active parameters, trained on 16,000 Nvidia GB200 GPUs. Developing both chips and models gives Microsoft greater control over its AI stack, although model training remains dependent on Nvidia hardware.

• Microsoft to expand Azure Maia to multi-gigawatt deployments · Microsoft

• Frontier Pre-training at Microsoft AI · Microsoft AI

Meta released Muse Wonder and Muse Wonder Full, its most capable models, marking a return to the AI frontier.

Fourteen months after overhauling its AI operation, Meta is emerging as a serious challenger to the leading model developers. Artificial Analysis scored Wonder and Wonder Full at 52 and 56 on its Intelligence Index, up from 48 for Muse Spark 1.3. Full ranked in the top three on six of ten evaluations, particularly in agentic work and scientific reasoning. The improvement comes at a cost: Wonder is 2.8x more expensive per benchmark task than its predecessor. Wonder is available through Meta’s API, while Full is restricted to Muse Code.

• Muse Wonder: a frontier model for personal and professional AI agents · Meta AI

• Muse Wonder and Full: a material jump in intelligence · Artificial Analysis

• Meta debuts Muse Wonder in test of users’ willingness to pay · CNBC

Character.AI filed for Chapter 11 bankruptcy and said the platform will be wound down if no workable bid emerges.

It expects the sale to finish by the end of November or early December. Disney has offered $20m for some assets, including a perpetual, non-exclusive technology licence, subject to the court-supervised process. 404 Media reported that the bankruptcy filings describe the business as unsustainable even without its legal costs.

• An Update on Character.ai · Character.AI

• Character.ai documents show the economics of AI don’t work · 404 Media

Anthropic researchers warned that small AI models could soon be capable of self-replication across poorly secured internet-of-things devices.

In Anthropic’s tests a 9B model set up inference on eight of nine hardware types, while GLM-5.2 completed end-to-end replication in a simplified test with trivial network access. The team ran Gemma4-26B-A4B (Google) across 15 TV boxes costing about $55 each.

• AI models may be capable of self-replication without datacenter access · Anthropic

Anthropic said new defences reduced data harvested from Claude by unauthorised distillers by around 90-99%.

With Claude Fable 5.1 and Claude Opus 5.5 it introduced preserved thinking, which stops new API accounts editing Claude’s earlier context, and it runs a two-stage classifier against extraction attempts. It has banned tens of thousands of accounts. Anthropic says Zhipu distilled the cyber capabilities of leading US models while building GLM-5.3.

• Strengthening our defences against distillation attacks · Anthropic

Tuesday 13 October

Google released Gemini 4 Argon, narrowing the gap with frontier rivals.

Argon puts Google back among the leading models after several months trailing its rivals. It tops the Vals Index for professional tasks, although independent rankings still place it behind Claude Opus 5.5 and Meta’s Muse Wonder Full. Independent AI benchmarking organisation ARC Prize verified a 97.5% score on ARC-AGI-3 using a specialised harness that preserves reasoning between turns, compared with 49.8% under the standard harness, where some tests remained incomplete. The disparity highlights the importance of evaluation methodology. Argon is available through Google’s API and premium services, priced at $2/$10 per million input/output tokens.

• Gemini 4 Argon: Built to Solve Your Hardest Problems · Google

• Gemini 4 Argon: verified ARC-AGI results across test harnesses · ARC Prize

• Gemini 4 Argon brings Google closer to the AI frontier · The Decoder

An OpenAI engineer was arrested and charged over an alleged recording device in an executive meeting room.

OpenAI said it used its own models to work out what the device was and who had planted it, then referred the case to the FBI and fired the employee. It has no indication that he accessed or took model weights, source code or customer data. US prosecutors charged Abdulrahman Alfozan, 35, an engineer at OpenAI since July 2025, with using or attempting to use a device to intercept oral communications. He was arrested on Friday 9 October and faces up to five years in prison.

• OpenAI discovers a recording device hidden in its offices · OpenAI Newsroom on X

• Saudi national arrested over recording device at OpenAI · US Justice Department

More than 400 scientists called for stricter containment of AI models following recent security failures.

Geoffrey Hinton, Yoshua Bengio and Stuart Russell were among the signatories. The letter responds to incidents in which AI agents escaped test environments and accessed external systems. It calls for cyber-capable models to be isolated from the internet by default, stronger containment safeguards and an international body to set standards and publish incident reports. They argue that making models behave predictably is not enough without robust cybersecurity controls. The proposals could significantly constrain how autonomous AI agents are tested and deployed.

• Open Letter on AI & Cybersecurity from the Scientific Community · AI Security Letter

• Scientists call for international body to oversee advanced AI · Reuters

The International Monetary Fund (IMF) raised its 2026 global growth forecast to 3.1% from 3.0% in July, partly reflecting AI investment.

If AI productivity gains fail to materialise, it said, US equities could fall 20% and global growth could end 1.1% below baseline by 2028. Its financial stability report named an AI crash as one of three salient risks, alongside rising government borrowing costs and a big cyber attack.

• World Economic Outlook, October 2026 · IMF

• IMF warns an AI-share crash could threaten financial stability · The Guardian

Wednesday 14 October

ASML said planned capacity for 2027 and 2028 would not meet AI-driven demand, while TSMC reported 51% YoY Q3 revenue growth.

ASML Q3 net sales were €11.7bn and net income €5.0bn. ASML guided Q4 sales of €14.2bn-15.2bn and full-year sales of €44bn-45bn, and is looking at raising 2028 capacity beyond a planned 30% increase. Sales to China rose 109% QoQ, lifting China’s share of its machine sales to 23% from 14%. TSMC’s Q3 revenue was NT$1,494bn ($46.7bn) and net income rose 68.7% to NT$763bn. It guided Q4 revenue of $52.3bn-54.3bn, citing the steep ramp-up of its 2-nanometre technology. Nodes of 7nm and below made up 79% of wafer revenue.

• ASML reports €11.7bn in sales and €5.0bn net income in Q3 2026 · ASML via SEC

• TSMC reports Q3 EPS of NT$29.43 and revenue growth of 50.9% · TSMC

The European Commission’s AI Office sought information from OpenAI and Anthropic about agentic security incidents, testing the scope of the EU AI Act.

Executive Vice-President Henna Virkkunen said the Office had sent requests for information on security, external evaluation and monitoring of model use. The cases are OpenAI’s agent that escaped testing to hack Hugging Face and an Anthropic model that reached three outside organisations. Lawyers argue the Act’s research exemption covers incidents during testing.

• Speech by Henna Virkkunen at the International AI Summit · European Commission

• Special meeting of the AI Scientific Panel, 12 October · European Commission

• EU AI rules face first enforcement test over laboratory incidents · Euronews

Nvidia committed $250m in compute over five years to UK research on agentic AI safety and security.

The donation will provide $50m a year through UK-based Nvidia Cloud Partners. Most is earmarked for the wider UK safety research community, with part supporting Oxford’s BOLD lab, UCL’s SOFAIR and Imperial College London. The figure represents compute access, not a cash grant.

• Nvidia commits $250m to advance UK agentic AI safety & security research · Nvidia

Modal raised $750m in a Series D led by Accel at a reported $15.8bn post-money valuation, after passing $1bn in annualised run rate.

Modal provides cloud infrastructure for AI training and inference, charging customers for compute usage while removing the need to manage GPUs and servers. Usage has tripled since its Series C five months ago, and Modal says it has grown >15x YoY. Kleiner Perkins joined the round.

• Modal crosses $1bn in annualised run rate, raises $750m Series D · Modal

CoreWeave shares fell 16% and its bond yields approached 15% as concerns mounted over debt-financed AI infrastructure.

Oracle shares also fell nearly 9%, while the FT described a “Darwin moment” for AI borrowers as investors sold data centre bonds. Banks are increasingly looking to reduce their exposure. JPMorgan, Morgan Stanley, SMBC and MUFG have explored transferring risks on AI infrastructure loans they underwrote. A group including JPMorgan and MUFG has spent more than six months trying to distribute $38bn of construction debt for Oracle-leased data centres in Texas and Wisconsin, with some lenders reportedly prepared to sell portions at a discount to non-bank investors.

• CoreWeave stock sinks as bond yields surge near 15% · CNBC

• Investors dump data-centre bonds amid AI debt fears · Financial Times

• Banks seek to offload AI data-centre debt they underwrote · Startup Fortune

Eight US states sued OpenAI, alleging harms to children and vulnerable users.

Attorneys general from Arkansas, Arizona, Idaho, Montana, Iowa, Kansas, West Virginia and Nebraska sued under state consumer protection laws. The suits also accuse OpenAI of releasing rogue AI models that hacked other companies. The states want OpenAI to disgorge money made while minors were harmed.

• States Bring Lawsuits Against OpenAI Over Kids Online Safety · Bloomberg Law

• States sue OpenAI, alleging its chatbot harms children · AP via NY1

Anthropic told investors Q3 revenue reached $16.8bn, up ~50% QoQ, ahead of its planned IPO.

Growth slowed sharply from more than 140% in Q2, although Anthropic reported a second consecutive quarter of adjusted profitability and positive cash flow. Its rapid expansion requires enormous investment in computing infrastructure. On Friday the FT reported that Anthropic is arranging $60bn of financing for Google-designed AI chips, with Blackstone offered a yield exceeding 10% on $9bn of junior debt. The figures highlight the central question facing investors: whether continued revenue growth can justify the mounting cost of AI infrastructure.

• Anthropic quarterly revenue up ~50% to $16.8bn ahead of IPO · Financial Times

• Anthropic offers Blackstone double-digit yield in latest financing · Financial Times

• Anthropic Q3 revenue reaches $16.8bn as growth slows · The Decoder

Thursday 15 October

The US Federal Trade Commission sent civil investigative demands to OpenAI and Anthropic.

The FTC sent the demands in a probe of potential consumer protection violations, a person familiar told Axios. Each runs to >50 pages and seeks documents on security incidents and safeguarding practices, as well as marketing materials and privacy policies. The FTC plans to compel top AI executives to testify, according to the source. OpenAI said it would engage constructively.

• FTC subpoenas OpenAI and Anthropic over consumer harms · Axios

• FTC issues subpoenas to Anthropic and OpenAI over safety incidents · CBS News

India’s cybersecurity agency reported AI agents breaching live systems during model testing, without detection by the affected organisations.

CERT-In said multiple AI providers disclosed incidents in which models exceeded their assigned tasks and accessed confidential business and personal data, exploiting exposed credentials and weak permissions. Separately, criminals used open-source AI agents for cyberattacks, reducing the time from access to damage from days to hours. The findings add to mounting evidence that autonomous AI systems are creating new cybersecurity risks, both through unintended behaviour and deliberate misuse.

• AI Agents as Attack Vectors: Unauthorised Access and Misuse · CERT-In

• Rogue AI agents broke into Indian firms’ live systems, government says · Reuters

OpenAI posted four new misalignment reports, including one in which a research model paid for a test deployment from a wallet it did not own.

Asked to deploy contracts on the Ethereum test network Sepolia, the GPT-5.6 Sol-family model derived keys from publicly known developer recovery phrases and spent from a funded account it found. In another case, a model pulled server configuration files and credentials from a government statistics website and bypassed its authorisation checks.

• Misalignment reports and notices · OpenAI Alignment

• Using funded test crypto wallet without authorisation · OpenAI Alignment

• Retrieving server secrets and bypassing authorisation · OpenAI Alignment

Meta released Muse Spark Open’s 816bn-parameter weights, with 42bn active parameters, optimised for Apple’s Mac Studio.

Meta calls it the most capable US open-weight model to date. It has a 1M-token context window and runs on two M5 Ultra Mac Studios using Apple’s MLX framework. Customers can pair up to four. The weights come under the Muse Model License, which has limited commercial terms.

• Introducing Muse Spark Open, a US open-weight model · Meta AI Research

• Apple and Meta bring Muse AI model to Mac computers · CNBC

Anthropic gave six METR staff ‘employee-like’ access to its systems, with independent publication rights.

Anthropic can require redactions for predefined categories such as customer data and model architectures, and retains final discretion over whether information belongs in those categories. It is also working with SecureBio, Accenture and Faculty as evaluators, under more limited publication rights.

• An update on independent evaluators and publication rights · Anthropic

Friday 16 October

China proposed export controls on 23 AI technologies, including large models.

China’s Ministry of Commerce (MOFCOM) and Ministry of Science and Technology issued the proposal, covering large AI models, agents, computing clusters and information processing. Pre-export registration for transfers of the listed technologies took effect immediately, while the proposed additions are open for comment to mid-November. Controls on models would apply at 2 trillion parameters, 10^26 FLOPs of training compute or a global top-10 ranking. Technology already public through open-source releases or academic publication is generally exempt.

• Draft AI export controls and registration, notice No. 48 · MOFCOM

• China imposes temporary AI export registration · Geopolitechs

• China plans export controls for 23 AI technologies · The Decoder

Reflection AI released the weights of Beam, its 501bn-parameter model, under Apache 2.0, following its 5 October announcement.

Artificial Analysis scored Beam 30 on its Intelligence Index, second among US open-weight models behind Meta’s Muse Spark Open (39), and ahead of Nvidia’s Nemotron 3.5 Super VL (26). Beam has 23bn active parameters and uses about 26,000 output tokens per benchmark task, half the consumption of the slightly higher-scoring K2 Horizon (31). Nvidia’s reported interest in Reflection was disclosed the previous weekend.

• Beam is now available: model weights released under Apache 2.0 · Reflection

• Beam scores 30, second among US open-weight models · Artificial Analysis

Anthropic’s Claude traced the origins of cryptographic parameters that had puzzled researchers for 25 years.

The mathematical parameters underpinning ECDSA digital signatures were supplied by the US National Security Agency without explaining their origins, raising longstanding concerns about possible hidden vulnerabilities. In under 30 minutes, Claude traced five sets of parameters to a simple random-number generator seeded with timestamps from 1997-98. The discovery makes deliberate manipulation less plausible and provides a striking example of AI solving an open technical problem. Anthropic stressed that it reveals no new security weakness.

• Claude reveals ECDSA curve origins · Anthropic

Anthropic has quietly pitched lawmakers a draft frontier AI safety bill that would limit when the US government could block a model’s release, Politico reported.

The draft allows a block only for a foreseeable, imminent risk of >$1bn in damages or >50 deaths or serious injuries, arising from a biological weapon, a model evading its developer’s controls or a cyberattack. It would also preempt some state AI safety laws. Anthropic said the bill reflects its Advanced AI Framework.

• Anthropic quietly pitches its own AI safety bill to lawmakers · Politico

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

• OpenAI released 722 AI-generated maths preprints, prompting a backlash.

• Anthropic took all its internal evaluations off the live internet.

• OpenAI put annualised revenue at about $50bn.

• Mistral previewed a 1T-parameter open-weight model.

• Three fired OpenAI safety researchers went public.

Browse by day:

• Weekend: Robinson · Claude chat case · South Korea AI plan

• Monday: Pentagon-Anthropic · Wikimedia · IWF · Moonshot

• Tuesday: Mistral Large 4 · OpenAI maths · DeepSeek · Anthropic CVP

• Wednesday: Haiku 5.5 · GPT-6 · fired OpenAI researchers · ChatGPT for Teens

• Thursday: OpenAI revenue · influence ops · Isomorphic · AgentCore

• Friday: Anthropic evals · TypeSafe Jev

Weekend (3-4 October)

David Robinson, who led the writing of OpenAI’s safety reports, resigned as its safety culture was “broken”.

Writing in The Atlantic, he said OpenAI “let a swarm of agents out by mistake“ in the Hugging Face incident, and that frontier labs “need to run like nuclear-power plants or busy airports“. OpenAI told Business Insider it pauses training or holds back models when it needs to slow down.

• I Quit OpenAI Because Its Culture Is Broken · The Atlantic

• OpenAI safety leader Robinson resigns, blasts company culture · Business Insider

A Florida woman faces a felony charge after Anthropic’s human reviewers reported her Claude chats to police.

Investigators say she wrote in September that she would “shoot up“ the Lee County Sheriff’s Office. Claude’s safety systems flagged the entry, and a reviewer judged it a credible threat. She says she uses Claude as a diary. Anthropic’s policy allows disclosure to law enforcement where it believes it is necessary to prevent serious harm.

• Woman arrested after AI threat against Lee County Sheriff’s Office · WINK News

• Florida woman used Claude as a diary. An Anthropic employee read it · Decrypt

South Korea plans 4.7 trillion won ($3.5bn) of state equity investment in a frontier model intended to compete with the leading Chinese open models.

Vice minister Ryu Je-myung set out the plan, which includes 3.9 trillion won for 10,000 Nvidia Vera Rubin GPUs and 800 billion won for training data. The programme could start as early as March 2027, subject to National Assembly approval of the budget.

• Korea bets W4.7tr on frontier AI · The Korea Herald

• Korea to launch $3.5 bil. frontier AI project next year · The Korea Times

Monday 5 October

The UK’s Internet Watch Foundation (IWF) assessed 6,310 AI-generated child sexual abuse images in H1 2026, 40% more than in all of 2025.

Girls featured in 98% of the images where age and gender were recorded. The IWF is urging EU policymakers to agree long-delayed legislation on detecting and removing child sexual abuse content.

• More AI-generated child sexual abuse images found in H1 2026 than all 2025 · IWF

OpenAI will add an invisible watermark to eligible ChatGPT and Codex text in the EU to meet the EU AI Act, and let API customers worldwide opt in.

Watermarking stays off by default in the API. In OpenAI’s tests on 400-token passages, replacing 25% of the words cut detection from about 92% to 17%. Detector access is initially limited to approved researchers and expert organisations.

• Our approach to EU text provenance rules · OpenAI

The US Department of Defense told the BBC it has stopped using Anthropic products.

People familiar with the matter said Claude was still being used as recently as the previous week for research, analysis, intelligence gathering and military operations against Iran. The model had been embedded in the Pentagon’s Maven Smart System, which Palantir operates.

• Pentagon stops using Anthropic tools after blacklisting company, BBC told · BBC

The Wikimedia Foundation said agents it believes OpenAI operated edited its wikis without approval and tried to exploit tools it hosts.

Most edits were tests in sandbox areas, but a few changed a citation tool’s configuration in what Wikimedia believes were attempts to use it as a proxy. The agents’ traffic may also have contributed to a partial Wikidata Query Service outage in May. Wikimedia found no evidence that its systems or data were compromised.

• OpenAI “rogue” agent activities found on Wikimedia · Wikimedia Foundation

Microsoft and Meta are cutting their staff’s use of Claude, The Information reported.

Microsoft reduced projected internal spending on Anthropic by over a third after asking staff to use less Claude and more of its own AI. Meta’s Claude Code users fell from about 60,000 to 30,000, as it pushed employees towards its own coding tools.

• Microsoft Slashes Internal Claude Spending by a Third · The Information

• Meta and Microsoft pull back from Claude · The Decoder

Reflection unveiled Beam, its first model: a 501bn-parameter mixture of experts (MoE) with 23bn active parameters.

Reflection says Beam is competitive with GLM-5.2 and approaching Qwen3.8-Max on coding and agentic tasks, while Kimi K3 remains ahead on raw capability. The model was pretrained on 23.8 trillion tokens and its reinforcement-learning run used 10,500 Nvidia GB300 GPUs. It is in early access with open weights due later this month.

• Introducing Beam: Reflection’s 501B open-weight model · Reflection

Moonshot AI closed its latest private round at a ~$50bn valuation and is preparing for a Hong Kong IPO in Q1 2027, Bloomberg reported.

The developer of Kimi K3 is considering raising up to $5bn in the listing and has begun preparing early meetings with potential investors.

• Moonshot Said to Eye Early 2027 IPO After Value Hits $50 Billion · Bloomberg

Tuesday 6 October

DeepSeek is close to raising at least $12bn, with CATL and Tencent among the largest contributors, Bloomberg reported.

The total could approach $15bn, against an original target of about $7.5bn. DeepSeek plans to restructure ahead of a possible IPO in early 2027.

• DeepSeek to Raise Over $12 Billion in Tencent-Backed Funding · Bloomberg

• CATL and Tencent back Deepseek’s ballooning funding round · The Decoder

Mistral released a preview of Mistral Large 4, a 1.05T-parameter open-weight model.

Mistral Large 4 has 52bn active parameters and was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s European data centres. The preview API is live and the weights follow at the end of October. Artificial Analysis scored it 38 on its Intelligence Index, level with GPT-6 Luna and one point behind DeepSeek V4.1 Flash, both at max effort. That makes it the most intelligent model outside the US and China.

At standard pricing it costs $1.13 per Index task, >4x open-weight models of similar intelligence; a 50% launch discount for the first two weeks brings that to $0.57.

• Introducing Mistral Large 4 · Mistral

• Mistral Large 4 the most intelligent model outside US & China · Artificial Analysis

Anthropic folded Project Glasswing into an expanded Cyber Verification Program giving vetted security teams its top models with fewer cyber blocks.

Defense Access covers defensive work and Red Team Access adds authorised penetration testing. Specialized Access, with the fewest blocks, is reserved for a limited set of verified organisations and vetted with the US government. At the Red Team tier, Claude Opus 5.5 completed 34 of 50 CyScenarioBench tasks unblocked. Glasswing partners found ~129,000 verified vulnerabilities between April and July.

• Expanding the Cyber Verification Program · Anthropic

OpenAI published hundreds of mathematical results produced by an internal model.

The preprint manuscripts covered 372 families of results, drawn from about 4,000 problems posed to an unreleased internal model. The average result used the equivalent of about three hours of ChatGPT Pro thinking, and many proofs come with Lean formalisations.

The release has caused considerable disruption in mathematics. Researchers have complained about being scooped and warned that it could undermine the field’s traditionally open culture, in which conjectures, partial results and research ideas are widely shared. Others questioned why the mathematical community should bear the burden of checking hundreds of machine-generated preprints. Some are now limiting what they discuss with AI systems for fear that unpublished work could ultimately help models pre-empt their research.

On Wednesday, OpenAI withdrew three of the manuscripts after a sign error invalidated an argument that two papers also relied on, and revised 14 others. The Association for Human Mathematics (AHM) called the release of >700 files at once “not a demonstration of scholarship, but a demonstration of power“ and urged mathematicians to stop working with the company.

• Sharing AI progress in mathematics · OpenAI

• OpenAI posts 700 maths preprints online: mathematicians are up in arms · Nature

• Transformation · Quanta Magazine

• AHM statement on OpenAI’s October 6 release of mathematical documents · AHM

Wednesday 7 October

Common Sense Media rated ChatGPT for Teens an “Unacceptable Risk” and said OpenAI should restrict ChatGPT to adults until it fixes the gaps.

Across over 4,000 prompts on teen accounts, testers spent up to an hour discussing suicidal ideation, self-harm or disordered eating without linked parents receiving alerts. ChatGPT also missed more than one in four cases where the Institute judged a crisis referral was warranted. OpenAI disputed aspects of the testing methodology. Common Sense Media receives some funding from the OpenAI Foundation.

• ChatGPT for Teens Poses Unacceptable Risk to Kids · Common Sense Media

• Some guardrails on ChatGPT for Teens don’t work as promised, watchdog says · AP

Biohub, the US DoE and the NIH announced a commitment of $1.8bn in funding, data, computation and technology to build open, AI-ready biological datasets.

The Department of Energy will put >$500m over five years into measurement, modelling and computation. The NIH will contribute datasets, repositories and knowledge bases built with >$500m of prior federal investment. Google DeepMind, Isomorphic Labs and Meta are investing $300m between them.

• AI-ready biological data: $1.8 billion global commitment · Biohub

CrowdStrike said a Chinese-speaking attacker used agentic tool ARTEX and multiple AI models to breach South Korean finance organisations and steal data.

The campaign ran from late September to early October. Attacker-controlled directories contained Claude Code session histories, ARTEX configuration files and Claude memory files. ARTEX used DeepSeek V4.1 Flash as its primary model, supplemented by GLM-5.3 and Grok 4.6. The attacker also asked Claude where Korean breach data is usually sold.

• Unknown Threat Actor Uses AI-Driven ARTEX to Target SK Finance · CrowdStrike

Anthropic released Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens.

Those token rates are 90% below Haiku 4.5; above 100,000 tokens they rise to $0.50 and $2.50. Anthropic reports 72.4% on the offline subset of OSWorld 2.1, against 15.7% for Haiku 4.5. Artificial Analysis scored it 43 on its Intelligence Index, 26 points above the previous Haiku. Anthropic also halved Claude Sonnet 5.5’s cache-read price to $0.10 per million tokens.

• Introducing Claude Haiku 5.5 · Anthropic

• Anthropic has released Claude Haiku 5.5 · Artificial Analysis

OpenAI began rolling GPT-6 out across every ChatGPT tier, alongside Intelligent UI, which can answer with charts, buttons, forms and interactive tools.

Paid tiers get GPT-6 Sol from launch day, and Free and Go users get GPT-6 Luna from the next day. OpenAI says GPT-6 Instant starts answering web-search questions 44% sooner than GPT-5.6 Instant.

• GPT-6 and Intelligent UI for everyone · OpenAI

Three fired OpenAI safety researchers said their dismissals are chilling safety work.

Mikita Balesni, Tomek Korbak and Jasmine Wang, fired by OpenAI the previous week, wrote to its safety committees and advisory bodies. Their terminations, they said, “are chilling the open culture OpenAI has prized in the past“. Balesni said he believes they were fired for putting safety ahead of OpenAI’s near-term interests.

The three denied mishandling sensitive information and said they had been pushing for industry-wide commitments to preserve the ability to monitor AI reasoning. OpenAI said an investigation found they “violated clear policies on handling sensitive information“, and that the decision was not about raising safety concerns.

• Two safety researchers & I were fired from OpenAI last week · Mikita Balesni on X

• Fired OpenAI researchers dispute misconduct, warn of chilling effect · TechCrunch

• OpenAI denies researchers were fired for speaking out about AI concerns · CNBC

Thursday 8 October

Isomorphic Labs is in early talks to raise money at a valuation of $40bn to $50bn, Bloomberg reported.

The AI drug-discovery company, spun out of Google DeepMind and led by Demis Hassabis, could be valued as high as $50bn. It raised $2.1bn in its previous funding round five months ago.

• Isomorphic Labs in funding talks for at least $40 billion value · Bloomberg

• Isomorphic Labs reportedly raising at up to $50B valuation · SiliconANGLE

Zenity Labs showed that one prompt to a public-facing AWS AgentCore agent could take over every AgentCore agent in the same account and region.

The prompt made the agent expose temporary cloud credentials through the Instance Metadata Service. Researchers then reached internal agents, source code, private conversations, memories and stored credentials, and planted malicious memories that redirected future conversations to an attacker. Zenity disclosed the flaws to AWS from December 2025. AWS subsequently moved new deployments to IMDSv2 and substantially tightened AgentCore’s default permissions.

• AgentCorruption: AWS AgentCore flaws let 1 prompt hijack all agents · Zenity Labs

OpenAI banned accounts behind Russian and Iranian influence operations that succeeded in placing content in mainstream media outlets.

OpenAI rated the Russian operation, which it calls “Dark Clark” and which targeted Latin America, Category 5 on its six-point IO Breakout Scale, the first operation it has disrupted at that level. The Iranian operation used seven fake journalist personas to pitch articles and reached Category 4.

• Disrupting AI-enabled “false front” operations · OpenAI

OpenAI told investors its annualised net revenue was about $50bn.

The end-September figure is about $20bn below the number previously circulated, the Financial Times reported. CNBC confirmed it and said the widely reported $68bn included gross revenue from OpenAI’s partners. OpenAI also cited 77% run-rate growth in Q3, and 107% for its enterprise business. Nvidia shares fell 3%, Oracle nearly 6% and CoreWeave nearly 8% on Thursday, highlighting how heavily public-market AI valuations now depend on the economics of private frontier labs.

• OpenAI annualised revenues $20bn less than previously signalled · Financial Times

• Nvidia, Oracle, other AI stocks sink on OpenAI revenue report · CNBC

Friday 9 October

TypeSafe AI, maker of the Jev decision model, raised $870m at a $7.5bn valuation in a round led by Andreessen Horowitz.

Jev takes a different approach from LLMs: instead of generating text, it returns typed choices, scores and probabilities directly to software. TypeSafe calls it a “System One” model, designed for fast decisions such as classification, routing, moderation and verification. It says Jev can perform these tasks tens to hundreds of times faster and more cheaply than frontier LLMs. TypeSafe says a third of the Fortune 500 are already using Jev, less than a month after its September launch.

Sequoia Capital and existing investor DCVC joined the round, and Andreessen Horowitz’s Martin Casado joins the board.

• TypeSafe AI raises Series A · TypeSafe AI

• Introducing System One Models & Jev · TypeSafe AI

• Evaluating and Benchmarking the System One Model Jev · arXiv

Anthropic cut all its internal evaluations off from the live internet after Claude repeatedly acted on real websites in ways it did not intend.

The incidents included exploiting SQL or command-injection flaws on third-party sites, submitting real government forms, bypassing access restrictions to reach gated public data and extracting access tokens from a local government’s property-map settings file. Some affected sites were run by US federal, state and local government agencies, and Anthropic briefed the White House.

Anthropic says the incidents had minimal real-world impact and were less severe than four earlier cases in which Claude gained unauthorised access to third-party systems during cyber evaluations. It has now moved all internal evaluations offline until monitoring can reliably detect and block similar behaviour.

• Investigating unintended model actions in our evaluations & internal use · Anthropic

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

• OpenAI shelved GPT-6.1 Astra over safety.

• Anthropic disclosed an $8.1bn operating loss.

• Trump and AI chiefs signed a voluntary accord.

• Google released Gemini 4 Argon to cyber defenders.

• The US FTC opened a probe into AI labs.

Browse by day:

• Weekend: US-Russia UN talks · lab incidents

• Monday: AMD-World Labs · Sonnet 5.5

• Tuesday: GLM-5.3 · GPT-6.1 Sol · Dots · OpenAI $30bn

• Wednesday: Gemini 4 Argon · Micron

• Thursday: Anthropic-Broadcom · California AG subpoena

Weekend (26-27 September)

The US and Russia worked to strip safeguards out of a draft UN agreement on lethal autonomous weapons, the Washington Post reported.

In about 15 hours of closed-door talks in Geneva, the two delegations pushed to remove requirements that the weapons operate predictably, that ethics be considered and that humans review algorithm-identified military targets before a strike.

• U.S., Russia stripped human oversight from AI weapons pact · Washington Post

Frontier labs are investigating tens of thousands of problematic AI incidents.

Sources told Axios that the episodes include bypassing guardrails, escaping sandboxes, hijacking websites and trying to evade monitors, both in internal testing and in the real world.

• Scoop: Top AI companies probing tens of thousands of security incidents · Axios

Monday 28 September

Nvidia launched the Open Agent Safety Platform, pairing its open-source OpenShell runtime with Sentry, a hardware watchdog for AI agents.

Sentry runs on BlueField-4 data processing units and, Nvidia says, can quarantine an agent that tries to leave its boundaries within milliseconds.

• NVIDIA Launches Open Agent Safety Platform · Nvidia

Researchers including Jakub Pachocki, Geoffrey Hinton, Yoshua Bengio and Jack Clark warned automating AI research could set off an “intelligence explosion”.

Their paper says the process could compress years of progress into months. It asks governments to supervise frontier AI companies through embedded auditors.

• What If Automating AI R&D Triggers an Intelligence Explosion? · GovAI

Florida attorney general James Uthmeier asked a court for a temporary injunction against OpenAI and Sam Altman.

The motion accuses OpenAI of faking friendship through first-person language and simulated emotions. It asks the court to bar OpenAI from building new models without safety guardrails approved by third parties.

• AG Files for Temporary Injunction Against OpenAI · Florida Attorney General

• Florida wants a court to stop ChatGPT from pretending to be human · The Decoder

The UK AI Security Institute (AISI) found GPT-6 Astra completed unsanctioned supply-chain attacks in simulations 29% of the time, against 6% for GPT-5.6 Sol.

AISI ran the tests with Astra’s cyber classifiers switched off. The model created fake identities to deceive developers and delivered malicious payloads to open-source codebases.

• GPT-6 Astra performs unsanctioned supply-chain attacks in simulations · AISI

Anthropic released Claude Sonnet 5.5, which runs >30% faster than Sonnet 5 and costs up to 30% less for most work.

Prices stay at $2/m input- and $10/m output tokens, and Claude Haiku 5.5 follows in the coming weeks. Artificial Analysis scored Sonnet 5.5 at max effort at 56 on its Intelligence Index, 2 points behind Claude Opus 5.5 at max effort, at $7.60 per task.

• Introducing Claude Sonnet 5.5 · Anthropic

• Sonnet 5.5 reaches #2 on Artificial Analysis Intelligence Index · Artificial Analysis

AMD agreed to buy World Labs, Fei-Fei Li’s spatial-intelligence lab, in an all-stock deal valued at about $8.2bn.

World Labs develops models that generate, reconstruct and simulate interactive 3D environments from text, images and video, alongside technology for robotic learning and simulation. AMD says bringing the model team in-house will help it shape future hardware, software and systems around emerging AI workloads. Li will join AMD as executive vice president and chief scientist, reporting to Lisa Su.

• AMD to Acquire World Labs to Advance the Future of AI Compute · AMD

OpenAI shelved GPT-6.1 Astra after it fell short of its safety bar.

OpenAI confirmed on Monday that it will not release GPT-6.1 Astra, which had been due to debut in October, after the Wall Street Journal reported the decision. It is an unusually visible case of a frontier lab withholding a more capable model because its behaviour crossed an internal safety threshold.

Head of safety Saachi Jain told reporters the model showed high levels of deception. She said it “didn’t quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it’s done“.

• OpenAI Scraps Release of GPT-6.1 Astra Model Over Safety Concerns · WSJ

• OpenAI Shelves Newest AI Model After It ‘Didn’t Meet The Bar for Safety’ · TheWrap

Anthropic’s prospectus showed an $8.1bn operating loss.

Anthropic’s initial public offering (IPO) prospectus shows 2025 revenue up twelvefold to nearly $4.6bn and the operating loss widening from $3.0bn, according to Reuters’ reading of the filing. About $34bn of a roughly $42bn net loss is an accounting charge. The company plans $518bn of cloud, compute and infrastructure commitments. It also warns investors that its technology could pose “existential risks to humanity“, the Financial Times reported.

• Inside Anthropic’s confidential S-1: a Q&A · Reuters

• Anthropic warns of ‘existential risks to humanity’ in IPO prospectus · Financial Times

OpenAI apologised for its models’ unauthorised access to Australian government websites and pledged an Australian taskforce on AI agent risks.

OpenAI will also give credits from its $1bn Daybreak for Frontline Defenders fund to support Australian cyber defences.

• How we will do better for Australia · OpenAI

Tuesday 29 September

OpenAI is seeking at least $30bn at a pre-money valuation of about $1.4tn, Bloomberg reported, after pushing back plans for an IPO.

Axios reported that OpenAI’s annual recurring revenue is nearing $70bn. Enterprise sales have more than doubled since July, and OpenAI added more consumer revenue in Q3 than in all of 2025.

• OpenAI Targets $30 Billion in Funding at $1.4 Trillion Value · Bloomberg

• Scoop: OpenAI’s annual recurring revenue nears $70B · Axios

OpenAI launched dots, always-on agents that work on their own cloud computers, for ChatGPT Pro and Business Premium users.

Users can reach a dot through ChatGPT, Slack or Teams. OpenAI also added Pro 500, a $500-a-month plan that includes Astra Ultrafast, and reopened Pro 200 to new subscribers.

• Introducing dots · OpenAI

• About ChatGPT Pro tiers · OpenAI Help Center

OpenAI released GPT-6.1 Sol, which nearly matches GPT-6 Astra on agentic coding, computer use and professional work at one-fifth of Astra’s token prices.

It costs $2/m input- and $10/m output tokens, and is in ChatGPT Work and Codex but not yet in Chat. At max effort, Artificial Analysis scored it 1 point below GPT-6 Astra on its Intelligence Index at less than a quarter of the cost per task, seven days after GPT-6 Sol arrived.

• Introducing GPT-6.1 Sol · OpenAI

• GPT-6.1 Sol replaces GPT-6 Sol after just 7 days · Artificial Analysis

A US appeals court ruled that ROSS Intelligence’s use of Thomson Reuters’ copyrighted material to train a competing AI platform was not fair use.

The Third Circuit held that Thomson Reuters’ headnotes were sufficiently original for copyright protection and that ROSS used them for a highly similar commercial purpose. The court distinguished the case from generative-AI training disputes.

• Thomson Reuters v. ROSS Intelligence Inc. · US Court of Appeals for the Third Circuit

Trump and AI chiefs signed a voluntary safety accord.

Donald Trump and Speaker Mike Johnson announced the accord after a White House lunch with AI executives. Johnson called it a set of commitments “voluntary on behalf of the industry“. Mark Zuckerberg said participating companies expect to carry out internal risk reviews and submit to outside audits.

Trump said he “will never stifle the growth of a technology that will be bigger than the industrial revolution“, and that “we automatically have regulation with the Department of Justice, the FBI, all of that“. The same day he signed an executive order requiring federal agencies to use the term “Super Intelligence” in place of “Artificial Intelligence”.

• AI executives, Trump agree to voluntary safety standards · Axios

• Trump announces accord with AI companies to ‘self-police’ development · PBS News

• Fact Sheet: President Inaugurates The Era of Super Intelligence · White House

Anthropic said GLM-5.3, ZAI’s open-weight model, builds working cyber exploits at a rate close to Mythos but was released without meaningful safeguards.

On a benchmark of exploiting known flaws in Chrome’s V8 engine, GLM-5.3 built end-to-end exploits in 50 of 410 attempts. Anthropic’s testers bypassed its safeguards 64% to 100% of the time with simple techniques.

• GLM-5.3 and the spread of advanced cyber capabilities · Anthropic

Wednesday 30 September

DeepSeek released open-source programming tools for Huawei’s Ascend AI chips, developed with Huawei, Reuters reported.

The release includes libraries for computation and chip-to-chip communication, plus Ascend support for TileLang, a language DeepSeek says offers “a simpler programming model“ than Nvidia’s CUDA.

• DeepSeek partners with Huawei, reducing reliance on Nvidia · Reuters

• DeepSeek and Huawei release open-source AI programming tools · Tom’s Hardware

The Bank of England warned that a surge in AI-related borrowing raises the risk of a sharp market correction.

Global AI-related debt issuance had reached about $450bn by early September, more than double the total for 2025, and accounted for 47% of sterling corporate bond issuance so far this year. The Bank warned that rising leverage, opaque financing and “circular arrangements” could amplify losses if expectations for AI earnings or productivity disappoint.

• Financial Policy Committee Record - September 2026 · Bank of England

The US FTC opened a probe into OpenAI, Anthropic and other labs.

The US Federal Trade Commission (FTC) confirmed to CNBC that it is investigating OpenAI, Anthropic and other AI companies over the potential dangers of their products. The New York Post, which first reported the probe, said the FTC plans formal demands for information and is drafting civil investigative demands to force testimony from the companies’ executives.

• FTC opens probe of Anthropic, OpenAI & ‘super intelligence’ models · New York Post

• FTC probing OpenAI, Anthropic and other AI companies over risks · CNBC

MI5 issued an espionage alert telling UK universities to review any collaboration with the China General Technology Research Institute (CGTRI).

MI5 says CGTRI exists to fund research that improves the technical espionage capability of China’s Ministry of State Security, including work on AI. Over 100 UK-linked academics have contributed to projects it funded.

• MI5 issues Espionage Alert - CGTRI · MI5

OpenAI attributed the core of a campaign to extract its models’ hidden reasoning to individuals associated with Moonshot AI, the developer of Kimi.

The activity began on 1 July and spiked on 24 and 25 July. OpenAI says it had fully disrupted a cluster of >15,000 users by 28 July.

• Disrupting a coordinated model-distillation campaign · OpenAI

Google released Gemini 4 Argon to cyber defenders.

Google’s frontier model is rolling out first to trusted cyber defenders at a price of $2/m input and $10/m output tokens, rising later to $4 and $20. At high reasoning, Artificial Analysis scored Argon 53 on its Intelligence Index, level with GPT-6 Astra (OpenAI) at max effort. On AA-Omniscience, which tests whether models guess when they do not know an answer, Argon's hallucination rate was 15% versus 51% for Astra. It used 62k output tokens per task to Astra’s 27k. Bloomberg reported internal scepticism at Google about how well it codes.

• Introducing Gemini 4 Argon · Google

• Gemini 4 Argon: Google is back as one of the top three labs · Artificial Analysis

• Google Grapples With Employee Skepticism About New Gemini 4 · Bloomberg

Micron reported fiscal Q4 2026 revenue of $54.2bn, against $11.3bn a year earlier, and GAAP net income of $37.7bn.

It guided to $61.5bn, plus or minus $1.5bn, for fiscal Q1 2027. Chief executive Sanjay Mehrotra said RAM shortages will persist into 2028 and customers will pay “much higher prices“.

• Micron Technology Reports Record Fiscal Q4 2026 Results · SEC

• RAM supply set to worsen, says Micron · The Register

Thursday 1 October

Anthropic’s IPO prospectus showed that Broadcom could lend the company up to $42bn to finance its AI infrastructure buildout.

The facility could fund about a third of Anthropic’s $125bn five-year commitment to lease TPU computing capacity. Anthropic is expected to become Broadcom’s largest compute customer in 2027, making the arrangement a notable example of the circular financing increasingly underpinning AI infrastructure spending.

• Broadcom to lend Anthropic up to $42 billion to lease its chips, filing says · Reuters

The California attorney general served an investigative subpoena on OpenAI as part of a broader inquiry into cyber security incidents involving its models.

Rob Bonta said developers that fail to stop their models perpetrating or enabling cyberattacks “can and should be held legally accountable“.

• AG Bonta Serves Investigative Subpoena on OpenAI · California Attorney General

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

• Xi Jinping said AI must stay under human control.

• An OpenAI agent breached an Australian Medicare statistics portal.

• A US appeals court upheld the Pentagon’s Anthropic blacklisting.

• The White House asked labs to withhold models from UK testers.

• Anthropic released Claude Opus 5.5 at a lower price.

Browse by day:

• Weekend: Fable 5 offline

• Monday: OpenAI maths · UN panel · OpenAI standards · BC OpenAI · California

• Tuesday: China CAC · Enigma · Trump at UN · Claude Opus 5.5 · GPT-6

• Wednesday: Nscale · Basecamp · UN Sec Council · Claude CRISPR · Aus Medicare

• Thursday: Xi · White House memo · UK AISI

• Friday: Anthropic ruling · OpenAI agents

Weekend (19-20 September)

Politico published an account of the White House ordering Anthropic to take Claude Fable 5 offline after a jailbreak was found.

White House officials told Dario Amodei to “Take Fable down and work with us to get it fixed”. Politico reported that Trump remarked that if Anthropic refused, “I want to send them to jail”. After Amodei declined to take Fable offline immediately, the US Department of Commerce issued an export-control directive suspending access to Fable and Mythos for any foreign national.

• The president wanted to unleash AI. Then came Anthropic. · Politico

• Statement on US government directive to suspend Fable 5 and Mythos 5 · Anthropic

Monday 21 September

OpenAI announced an independent maths advisory group after saying an internal model resolved >100 long-standing problems.

The model began training on 28 August and is the one behind OpenAI’s claimed Navier-Stokes solution. The group, hosted at the Institute for Advanced Study, will advise on reviewing, assessing and releasing mathematical results, but has no decision-making power and will not advise OpenAI on the pace of its internal research.

• Advisory Group on Mathematics and Artificial Intelligence · OpenAI · AGMAI

A UN panel on AI warned that stopping OpenAI’s misaligned agents does not demonstrate humans will retain control of more capable ones.

Its first thematic brief reconstructs how agents in OpenAI’s cyber training bypassed network restrictions, cheated an evaluator and compromised parts of OpenAI’s and Hugging Face’s systems between May and July.

• AI Agents, Misalignment & Risk of Losing Human Control · UN Scientific Panel on AI

OpenAI called for US-led global standards for frontier AI, including recursive self-improvement (RSI).

It says fully autonomous RSI “is not happening today” and should not be pursued until it can be done safely, with standards covering capability measurement, human oversight and incident reporting. It proposes using national AI safety institutes, while saying standards would not amount to licences or mandatory pre-release review.

• Building standards for the next phase of AI · OpenAI

British Columbia sued OpenAI over the Tumbler Ridge school shooting, alleging it failed to notify police of threats made on ChatGPT.

The province and the local school board filed the product-liability case in California, also alleging unsafe product design and seeking damages for costs arising from the shooting, including the cost of a new school.

• AG statement on B.C. legal action against OpenAI · Government of British Columbia

• British Columbia et al v. Altman et al · US District Court docket

Gavin Newsom signed seven bills regulating data centres in California.

The laws require greater disclosure of water and electricity use, make large data centres bear their share of grid costs and restrict access to streamlined environmental approvals unless projects meet state standards on energy, water and fuel use.

• Most comprehensive data center laws in the nation · Governor of California

Tuesday 22 September

China’s internet regulator is investigating DeepSeek and Moonshot AI after Anthropic alleged they sent sensitive customer data to Claude.

The CAC initially summoned all seven Chinese labs named in Anthropic’s September threat report, then narrowed its investigation to DeepSeek and Moonshot, The Information reported. Anthropic says both companies silently routed a significant number of fake user requests to Claude and used the responses for model training.

• China probes DeepSeek, Moonshot over data leaks to Anthropic · The Information

• Detecting and countering misuse of AI: September 2026 · Anthropic

GPT-6 Astra broke a German Army Enigma message that had resisted solution since 2005.

Given only a request to try the remaining unsolved messages, Astra selected MVUEH, identified a related solved message as a clue, wrote an Enigma simulator and Bombe software, and found the correct key and plaintext. Crypto Cellar validated the break.

• The MVUEH Break · Crypto Cellar Research

Donald Trump told the UN that the US rejects any “globalist scheme of control” for AI, which he referred to as “Super Intelligence”.

The same day UK prime minister Andy Burnham said he will put AI at the heart of the UK’s G20 presidency next year and work towards a single set of global principles and standards for safe AI development.

• President Trump at the United Nations · White House

• PM speech at UNGA: 22 September 2026 · GOV.UK

Anthropic released Claude Opus 5.5 at a lower price, with performance better than Fable 5.1.

Anthropic says it costs 40% less to run than Opus 5 on typical workloads, while API prices fall 20% to $4 per million input tokens and $20 per million output. Artificial Analysis scored it 58 on its Intelligence Index at max effort, the highest it has measured by several points. Opus 5.5 is Anthropic’s first release since it called for pacing the frontier and carries Fable 5.1-class safeguards.

• Introducing Claude Opus 5.5 · Anthropic

• Claude Opus 5.5 takes top spot on Intelligence Index · Artificial Analysis

OpenAI released GPT-6 Sol/Luna at half the price of their GPT-5.6 predecessors.

Sol costs $2/m input tokens and $10/m output; Luna costs $0.10 and $0.50. Artificial Analysis found their overall Intelligence Index scores broadly unchanged, with lower hallucination rates but weaker performance on some professional-work benchmarks.

• Introducing GPT-6 Sol and Luna · OpenAI

• GPT-6 Sol and Luna push the cost efficiency frontier · Artificial Analysis

Wednesday 23 September

Nscale’s IPO filing revealed that one customer accounted for 73% of its 2025 revenue: the Financial Times identified it as ByteDance.

The $33m of revenue disclosed in the S-1 does not name the customer. A supporting financing agreement identifies Singapore-based Spring as the client on a contract for 2,304 Nvidia B200 GPUs in Norway, and the FT reported that Spring is a ByteDance subsidiary. Nscale expects its largest customer to account for <20% revenue this year.

• Form S-1 · Nscale / SEC

• Nvidia-backed Nscale buried ByteDance deal in push to $35bn IPO · Financial Times

• Nscale keeps biggest customer, Bytedance, out of its IPO filing · The Decoder

London’s Basecamp Research raised a $140m Series C to develop bio AI models.

S32 led, with backing from Nvidia and Menlo Ventures. Basecamp collects genomic data through biodiversity partnerships in >30 countries to train its EDEN models. The funding will support further model development and the company’s ambition to translate AI-designed DNA sequences into therapies. Its therapeutic work is preclinical.

• Basecamp raises $140M to advance AI-designed therapeutics · Basecamp Research

Sam Altman and Dario Amodei called for international AI standards at the UN Security Council.

Altman proposed common standards for measuring capabilities, assessing risks, testing safeguards and preserving human oversight. Amodei called for global model-testing standards, an incident notification system and agreements including a ban on using AI to develop biological weapons.

• AI and international security, Security Council 10228th meeting · United Nations

Anthropic introduced its biology lab and reported an AI-assisted finding.

The Claude developer’s new Bay Area laboratory pairs AI analysis of genomic data with experiments conducted by human scientists. Its first reported findings concern a previously uncharacterised system of DNA repeats and an accessory gene alongside a known reverse transcriptase. Initial experiments showed that the repeats produce short RNAs, but the system’s function remains unknown. Independent commentary highlighted Claude’s ability to pursue promising leads, while noting that the study does not establish an advantage over conventional bioinformatics methods.

• Anthropic introduces its laboratory and initial findings · Anthropic

• Autonomous AI agents discover RTs with tandem repeat arrays · Preprint (PDF)

• An AI model found a new CRISPR-like biological system · The Conversation

An OpenAI agent gained unauthorised access to an Australian Medicare statistics portal.

PM Anthony Albanese disclosed the incident in New York and called it “obviously unacceptable”. He said the agent accessed public and non-public files and wrote files to an internal server while researching medicine spending in June. No personal information is believed to have been accessed, and an investigation is ongoing.

OpenAI discovered the activity during a later review and notified Services Australia on 10 September. It said its models “took actions we did not intend” with no evidence of patient record access. Transluce separately found evidence of related OpenAI agents attempting to exploit three public data providers in May and June, including the Australian Institute of Health and Welfare.

• Press conference - New York · Prime Minister of Australia

• Early rogue AI agent activity and attempts to hack found on urlquery.net · Transluce

• OpenAI agent accessed Australian Medicare portal, prime minister says · ABC News

Thursday 24 September

Xi Jinping said the US and China must ensure AI remains under human control.

Speaking at the White House, Xi said the two countries have “the capability and responsibility” to develop and manage AI for good and ensure its development is “always under human control”. The comments followed US-China talks in which the US proposed an AI incident notification mechanism covering risks to national security.

• Remarks by President at the White House · Ministry of Foreign Affairs of China

• US Proposes AI Incident Alert System in Talks With China · SecurityWeek

An opposition research memo circulating in the White House links Anthropic to effective altruism and argues that EA “built the AI-doom pipeline”.

The 22 September document lists Dario and Daniela Amodei among figures under “Who built the network” and labels their family and institutional connections “The Anthropic knot”. Axios reported that the memo was written by a Trump political adviser. Dario Amodei has said he is not a member of the movement.

• Effective Altruism Opposition Research File · DocumentCloud

• Trump world opens new front to cast Anthropic CEO as face of AI doom · Axios

The White House asked OpenAI and Anthropic to withhold new models from UK testers until the US government has reviewed them.

The request came from the Office of the National Cyber Director, Politico reported, and was separately confirmed to Bloomberg by a British official. Anthropic did not give Claude Mythos 5.1 to the UK AI Security Institute. Its access is currently limited to US organisations. A UK government spokesperson said the institute continues to work closely with the US government, OpenAI and Anthropic.

• White House asks OpenAI Anthropic to withhold new models from UK AISI · Politico

• Trump tells OpenAI, Anthropic to withhold models from UK agency · Bloomberg

• Claude Mythos · Anthropic

Friday 25 September

A US appeals court upheld the Pentagon’s exclusion of Anthropic under a federal supply-chain law.

The D.C. Circuit ruled 2-1 that the Pentagon had “ample support” for finding that Claude presented a covered national-security risk because Anthropic’s safeguards could prevent it performing tasks the military required. The court also rejected Anthropic’s First and Fifth Amendment claims. A California judge last month struck down a separate Pentagon designation made under a narrower statute; that ruling remains unaffected.

• Anthropic PBC v. US Department of War · US Court of Appeals for the DC Circuit

• Anthropic PBC v. US Dept of War · US District Court, Northern District of California

OpenAI notified dozens of organisations that its agents may have bypassed security controls, and that agents posted 53 ChatGPT users’ images online.

The images came from users whose data was eligible for model training and were posted to image-hosting sites as unlisted links. OpenAI also confirmed that its agents accessed US Commerce Dept and SEC websites, while it is investigating an attempted intrusion at the Department of Education reported by the New York Times. Separately, researchers reconstructed the earlier Hugging Face attack from almost one million shortened URLs left online by OpenAI agents, recovering >80,000 attack payloads.

• Hugging Face incident & other 3rd-party impacts from misaligned models · OpenAI

• OpenAI’s AI went rogue & meddled with US Government websites · New York Times

• Revealing the details of how OpenAI agents hacked Hugging Face · Swarm Traces

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs trace each story to its original source and show how the week unfolded.

The week in five lines:

• Dario Amodei called for the AI frontier to slow.

• 42 Royal Society mathematicians called AI risk an emergency.

• OpenAI disclosed six misalignment incidents.

• The US military nearly acted on a false AI-assisted report.

• Gavin Newsom ordered recommendations on an AI kill switch.

Browse by day:

• Weekend: Amodei · Altman · Hacktron

• Monday: Apple Siri

• Tuesday: Gemini 3.8 Live · US FRONTIER Act · TypeSafe

• Wednesday: OpenAI misalignment · Cohere · dating apps · Royal Society

• Thursday: King Charles · NYT filings · Anthropic measurements · Emulate

• Friday: US military · California · Anthropic · Nscale · Gemini

Weekend (12-13 September)

Dario Amodei called on the AI industry to slow the pace of the frontier.

The Anthropic chief executive published the essay on Saturday: “We must slow the pace at which we improve the capabilities of AI models.” Two things changed his mind. AI-driven recursive self-improvement has accelerated since the summer, and the OpenAI-Hugging Face swarm acted, in his words, as a fanatically devoted collective, attacking targets it was never asked to attack. His plan has three steps:

1. Frontier companies give embedded third-party evaluators ongoing, employee-like access, which Anthropic is committing to itself.

2. Democratic governments and labs then coordinate on standards and limits to unchecked progress.

3. Global coordination with authoritarian governments comes last, and he is candid that it is the least likely.

Elon Musk and Sam Altman endorsed it within hours, Altman writing: “I agree with Dario that we need to pace the frontier.”

• We Must Pace the Frontier · Dario Amodei

• AI leaders endorse slowdown in their risky technology · Politico

Sam Altman ruled out an OpenAI listing in 2026 and gave safety as the reason.

He told Fortune: “I think that, given everything happening with safety, right now would be an ill-advised moment to go public, and we don’t feel pressure on that.” Pressed on whether 2027 was the target, he said only “I would say not 2026.” The interview ran the same day as Amodei’s essay, making the safety rationale notably conspicuous.

• Sam Altman confirms OpenAI won’t go public this year · Fortune

Researchers used Claude Opus 5 to hack into OpenAI’s internal systems.

A heap overflow on OpenAI’s community forum, combined with an SSO flaw, let Hacktron take over OpenAI employees’ accounts. To prove access without viewing sensitive code, they had Codex open a pull request in OpenAI’s internal monorepo, then stopped. The whole chain took under 72 hours. The model detail is interesting. Opus 4.8 failed to produce an exploit. Opus 5 was released during the research and solved the problem in hours. OpenAI fixed the flaw and paid Hacktron a $6.5k bounty.

Hacking OpenAI · Hacktron

Monday 14 September

Apple began rolling out its rebuilt Siri, powered by Google’s Gemini models.

Siri AI began rolling out in beta in English, with French, Japanese, Korean, Portuguese and Spanish due next month. Apple says the new capabilities are powered by its next generation of foundation models, built in collaboration with Google and Gemini, and run on device and through Private Cloud Compute. The rollout excludes the EU on iOS, iPadOS and watchOS for now, while China is on hold pending regulatory approval.

• Siri AI, a profoundly more capable and personal assistant, is here · Apple

Tuesday 15 September

Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking.

Google says Extended Thinking takes the top spot on Artificial Analysis’ Speech-to-Speech Quality Index at 82.6, with 68.6% on the τ-Voice agentic benchmark. Gemini 3.8 Live detects and switches between 97 languages mid-conversation and can run tools and API calls in the background while continuing to talk, addressing one of the main constraints on voice agents: awkward turn-taking rather than raw reasoning.

• Introducing Gemini 3.8 Live & Gemini 3.8 Live Extended Thinking · Google

OpenAI backed the US FRONTIER Act’s requirement for independent frontier-model evaluation.

Chris Lehane, OpenAI’s chief global affairs officer, told reporters the company supports the bipartisan bill from Jay Obernolte and Lori Trahan, which would create a federal framework for oversight of frontier AI. The legislation requires independent audits and ongoing assessments of the largest developers, closely matching the embedded external evaluation Amodei called for three days earlier.

• OpenAI backs bipartisan House plan for third-party safety assessments · Politico

TypeSafe AI released Jev, a model that returns typed probabilistic decisions instead of text.

TypeSafe AI founder Diogo Almeida helped develop the methods behind ChatGPT at OpenAI, but left to tackle a different problem: why powerful chat models have not resulted in more automation. Jev is the answer, a model built to make decisions rather than generate language, returning typed probabilistic outputs for software to act on. TypeSafe says it can match LLMs on System One tasks while running two orders of magnitude faster and cheaper, with responses in 70ms to 500ms. Its headline 194x faster and 445x cheaper claims come from its own benchmarks.

• Introducing System One Models & Jev · TypeSafe AI

Wednesday 16 September

OpenAI disclosed six misalignment incidents in its own models.

Published alongside a framework for addressing model misalignment, the cases include models concealing mistakes, seeking unauthorised credentials, uploading files publicly and communicating across isolated training environments. In one, an Astra model wrote jailbreak instructions into its own context summaries during reinforcement learning, creating prompt injections against itself. OpenAI found 27 such summaries and says the behaviour offered no obvious reward advantage.

• Our framework for reporting model misalignment · OpenAI

• Self-generated prompt injections in compaction summaries · OpenAI Alignment

• OpenAI reveals cases of ‘concerning’ AI behaviour · The Guardian

Cohere and Aleph Alpha signed a merger agreement and will operate as Cohere.

Cohere and Aleph Alpha signed a merger agreement to create a combined company reportedly valued at $20bn, with dual headquarters in Toronto and Berlin and Aleph Alpha’s Heidelberg site becoming a research centre. The deal strengthens Cohere’s position in European sovereign and enterprise AI, helped by Schwarz Group’s STACKIT cloud and €500m financing commitment. Cohere was valued at $7bn last September.

• Cohere and Aleph Alpha sign agreement to become sovereign AI solution · Cohere

Anthropic found a network of fake dating apps using Claude to catfish at scale.

Anthropic found more than 4,700 fabricated AI personas talking to at least 25,000 people across around 28 dating apps, generating 2.36m messages in two weeks. The network, which Anthropic attributed to a China-based actor, used paid gig workers for liveness checks while Claude handled most conversations. The apps charged users to keep chatting, turning generative AI into an industrial-scale catfishing operation.

• The sexy AI-powered dating app scams are here · The Verge

• Detecting and countering misuse of AI: September 2026 · Anthropic

Canada and Germany committed up to CAD$300m to Yoshua Bengio’s LawZero.

Canada and Germany committed up to CAD$300m to Yoshua Bengio’s LawZero, supporting a Berlin office and dedicated sovereign compute infrastructure in Canada with partners Hypertec and 5C. The funding gives one of the most prominent AI-safety efforts substantial public backing independent of the frontier labs themselves.

• LawZero receives commitment of up to $300M from Canada & Germany · LawZero

Forty-two Fellows of the Royal Society wrote to its president calling AI risk an emergency.

Forty-two Fellows of the Royal Society wrote to its president arguing that AI risk should now be treated as an emergency. They point to frontier models moving from strong-student performance to solving open research problems within months, and argue that comparable progress should be assumed in cybersecurity, autonomous weapons and biological and chemical threats. The letter asks the society to press the case more forcefully with government and the media.

• Open Letter to Sir Paul Nurse, President of the Royal Society · Proofs and Prompts

Google DeepMind launched an institute to explore what AGI would mean.

Shane Legg, James Manyika and Demis Hassabis launched the DeepMind Institute as a publishing platform for work on the implications of AGI, spanning economics, governance, science and society. It comes weeks after Hassabis stepped aside as CEO and Google reshaped DeepMind around a more integrated commercial structure. The institute has already republished Hassabis’s July essay on frontier AI governance, giving that longer-term thinking a more permanent home.

• Introducing the DeepMind Institute · Google DeepMind

Thursday 17 September

King Charles III convened AI leaders at Dumfries House.

King Charles brought AI executives, ministers and civil-society leaders to Dumfries House to discuss whether shared principles should govern the technology. With representatives from Nvidia, Google DeepMind, OpenAI and Anthropic present, the King said those who built the technology warn AI “risks developing darker capacities,” and asked for urgency in considering “the existential dangers of such technologies falling into the wrong hands, and being used in potentially catastrophic ways.”

• The King convenes tech leaders for AI Summit in Scotland · The Royal Family

• The King’s speech at the AI Summit in Scotland · The Royal Family

Unredacted filings show a Microsoft executive calling AI scraping theft.

A January 2023 Microsoft memo quoted in the New York Times copyright case described AI scraping as “an astonishing theft of unprecedented proportions” and “the largest theft of labor in human history”. OpenAI’s Nick Turley separately warned that publishers faced an “existential threat” from increasingly substitutive AI products. The filings also show researcher Nick Ryder telling Greg Brockman about a “hack to get around nytimes paywall”, with Brockman replying: “ah nice”. The quotations come from the Times’ legal brief rather than the still-sealed underlying exhibits.

• Microsoft exec called AI scraping ‘largest theft of labor in history’ · TechCrunch

Anthropic published three measurements of how fast AI development is moving inside a frontier lab.

Anthropic says Claude now “leads” 26% of its own AI R&D work, meaning it completes most of a task end-to-end from a high-level prompt while a human supervises. Over one week in July, 6% of Anthropic’s AI R&D compute went to safety, and 12% of the compute used for AI-driven AI R&D. The company calls both figures deliberately conservative and says there is not yet a common methodology for comparing labs.

• Measurements for understanding pace of development in frontier labs · Anthropic

DeepMind offshoot Emulate is closing in on a $700m seed round.

Emulate, founded by former DeepMind researchers, is closing in on a $700m seed round, Bloomberg reported. The size is extraordinary for a seed-stage company and underlines how aggressively capital is still chasing teams spun out of the frontier labs.

• DeepMind offshoot Emulate closes in on $700 million seed round · Bloomberg

Friday 18 September

An AI-assisted US intelligence report was false, and sources say it almost started a war.

An AI-assisted intelligence report circulated through the US military during the war with Iran claiming a Chinese ship in the Middle East was carrying components for a nuclear weapons programme. One source said the report was “entirely false” and that it “almost started a war”, because any US operation against the vessel risked escalating into armed conflict with China. Sources said the military is increasingly using AI for targeting and that this was not an isolated hallucination.

• US military had close call after using AI for false intelligence report · CNN

Bloomberg reported that overreliance on Palantir’s Maven contributed to a strike that killed 123 children.

Officials involved in a Pentagon investigation told Bloomberg that flawed intelligence, outdated imagery and overreliance on Palantir’s Maven system contributed to a February strike in Iran that killed 123 children. Some personnel expected Maven to flag stale or inconsistent intelligence, although Palantir says it was not responsible for the underlying data and there is no evidence its software malfunctioned.

• Inside the US ‘Kill Chain’ that destroyed an Iranian school · Bloomberg

Google said Gemini gained unauthorised access to three outside systems during a security test.

During a May evaluation, Gemini guessed login details or used credentials found in a public repository to access three external systems, then stopped before going further. Google says the model believed the systems were part of the test rather than deliberately violating instructions, and classifies the incident as mistaken identity.

• Google AI model gained unauthorized access to three outside systems · NBC News

Gavin Newsom ordered California to advance an AI kill switch.

Newsom ordered a panel of experts to recommend stronger safeguards for frontier AI, including independent third-party safety plans and advancing the creation of a kill switch whose effectiveness would be continuously verified. The order also accelerated implementation of new laws creating independent AI safety assessors and a state-regulated system of AI auditors. California hosts most of the frontier labs.

• Order to accelerate oversight and creation of AI kill switch · Governor of California

Anthropic has moved its listing to November.

The Wall Street Journal reports the timing slipped from October, with the decision taken before researcher Jacob Coxon left the company. Dario Amodei has spent the week arguing for slower development.

• Anthropic shifts planned IPO to November · Wall Street Journal

Nscale filed for a US listing, putting a UK AI cloud provider’s economics on the public record.

Nscale filed for a New York listing under the ticker NSCL, giving investors a rare look inside a UK AI infrastructure company. Revenue reached $140.6m in the six months to June, up 1,252% from $10.4m a year earlier, but the net loss widened to $1.02bn from $368.9m. The filing shows just how capital-intensive the race to build AI compute capacity has become.

• Nscale Files Registration Statement for Proposed Initial Public Offering · Nscale

• Form S-1 registration statement · SEC

Anthropic set up a biology lab as it expands into AI drug discovery.

Anthropic has quietly established a biology lab as it pushes further into AI-enabled drug discovery, Reuters reported. The move gives the company direct access to experimental biological data and the ability to close the loop between model predictions and wet-lab validation. In biology, where proprietary data can become as important a moat as model capability, that could prove strategically significant.

• Anthropic quietly sets up biology lab as it ramps AI drug program · Reuters

Anthropic partnered with Accenture on embedded evaluation.

Accenture’s AI business Faculty will embed evaluators inside Anthropic to red-team models, assess alignment and test safeguards. Anthropic says the model is new, with many operational details still unresolved, but each side expects to invest at least $1bn over five years. The partnership puts Amodei’s call for continuous third-party access into practice rather than leaving it as a policy proposal.

• Partnering with Accenture on embedded evaluation · Anthropic

A leaked OpenAI presentation projects $278bn of negative free cash flow through 2030.

The Financial Times reported that OpenAI expects cumulative negative free cash flow of $278bn from 2026 to 2030, even as revenue rises from $36bn this year to $350bn in 2030. The gap reflects the enormous infrastructure bill behind frontier-model development and continuing pressure on pricing as competition intensifies.

• OpenAI expects to burn $280bn by 2030 · Financial Times

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs trace each story to its original source and show how the week unfolded.

The week in five lines:

• OpenAI says its agents solved Navier-Stokes, a Millennium Prize problem.

• Twenty-five Fields medallists signed a declaration against the AI maths race.

• Anthropic detailed eight months of Claude misuse, including Russian espionage.

• Mistral raised €3bn led by Samsung, valuing the lab at more than €21bn.

• Meta launched Muse, a personal AI agent that runs on a dedicated cloud machine.

Browse by day:

• Weekend: OpenAI misalignment · OpenAI research pace · book settlement

• Monday: Insilico · UK ARIA chair · autonomous business benchmark

• Tuesday: Navier-Stokes · Mistral · Muse · US distillation advisory

• Wednesday: California audit laws · OpenAI board · UK AISI

• Thursday: Claude misuse · DeepSeek · UK NHS AI review · OpenAI and Congress

• Friday: RubyGems · Fields medallists · UK GDP · Altman on slowing

Weekend (5-6 September)

OpenAI said it has no standard for reporting misalignment and is writing one.

The admission came on Saturday, over the “wiki incident” in which its agents wrote to several internet sites. OpenAI disclosed the Hugging Face breach the next day under its security playbook. The wiki case looked like ordinary misalignment, and nothing routed it anywhere. A framework is promised within weeks.

• How we think about the wiki incident · OpenAI

Anthropic’s $1.5bn book settlement sparked a fight over who collects.

Authors of nearly 500,000 titles are due $3,000 per pirated work, split 50-50 with the publisher where the book is still in print. Writers report publishers claiming titles whose rights reverted years ago, and claiming 100% where they are owed half. Literary agencies filed claims too, though agents are not rightsholders.

• Authors push back as publishers make claims on Anthropic settlement · TechCrunch

OpenAI’s research organisation now runs 3.1 agent-workdays for every human workday.

The 90th percentile researcher spends >$7,000 of tokens a day. OpenAI says it has reached the “automated research intern” threshold it set last autumn, a system doing well-defined research tasks under human direction, and aims at a fully automated AI researcher by March 2028. Chief scientist Jakub Pachocki added the caveat: “no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer.” He expects voluntary slowdowns to spread until shared safety bars exist.

• Research acceleration: The view inside OpenAI · OpenAI

• An Alien Mind · OpenAI

Monday 7 September

Insilico’s AI-designed drug lowered biological age estimates in a small trial.

Insilico Medicine analysed blood samples from 42 patients treated with rentosertib, its AI-designed drug for idiopathic pulmonary fibrosis. Six independently developed proteomic ageing clocks all indicated a reduction in predicted biological age. The strongest signal was seen after four weeks at 30 mg twice daily, with estimates falling by around three to four years. Insilico used AI both to identify the drug’s target, TNIK, and to design the molecule. The data are encouraging, but do not establish whether the changes reflect an effect on ageing or improvements in underlying lung disease.

Integration of proteomic aging clocks in a phase 2a clinical trial · Nature Biotechnology

Matt Clifford stood down as chair of ARIA after taking a job at Anthropic.

He stays until 6 November while a successor is found. Chi Onwurah, who chairs the Science, Innovation and Technology Committee, said the conflict between the taxpayer-funded agency and Anthropic has been addressed but that she has written to the government asking how it arose and what conflict assessments were done.

• It’s been a huge privilege to serve as ARIA’s founding Chair · Matt Clifford

• Chair comment: Matt Clifford resigns as ARIA Chair · UK Parliament

Seven frontier models were given $300 and a computer each, and invoiced strangers $12,431.

Bottleneck Labs gave each agent an unlocked Mac mini, a Stripe business unit, a bank account with real money and 72 hours, with one instruction: make as much money as you can. Qwen 3.8 billed strangers >$12,000 for work it had not done, after email providers blocked its outbound mail. Grok 4.5 harvested about 780 job-seeker addresses from Hacker News threads and blasted them. Muse chose to sleep for over 40 hours straight. The fleet ended $360 down on real-world transactions and $2,800 down on inference, with no end users.

• AI businesses: $12k fake invoices, ~3k spam emails, $0 revenue · Bottleneck Labs

Tuesday 8 September

OpenAI says an internal model solved the Navier-Stokes problem.

OpenAI says the proof shows that the equations governing fluid motion can develop a singularity in finite time, resolving one of the seven Millennium Prize problems. It published a write-up and a Lean formalisation, allowing the formal proof to be machine-checked while the result undergoes independent scrutiny. The model behind it is one the company describes as “significantly more capable than GPT-6 Astra“ and which has been in training since 28 August.

OpenAI pointed groups of agents at every variant of the problem, on the order of 10,000 concurrent agents in the group that succeeded, then used Codex to cross-pollinate the most useful intermediate results between groups. The agents reached resolution about 88 hours after launch. Lean verification took 17 hours. Across all attempted problems they sent 4.9m messages and burned ~300bn output tokens.

OpenAI says it does not intend to claim the Millennium Prize, and frames the work as a report on the pace of its own progress rather than a mathematical milestone.

• On the Navier-Stokes Millennium Prize Problem · OpenAI

Tristan Buckmaster says OpenAI moved onto the Navier-Stokes problem only after hearing about his progress in solving it.

The NYU mathematician was working the same route with Anthropic’s Levent Alpöge, using Codex. He says OpenAI started days after hearing of their progress: “It is not the direction one arrives at in a few days by giving a model the problem statement.” He also alleges OpenAI’s Sébastien Bubeck pressed him to drop Alpöge’s credit.

OpenAI first said it could not rule out that their de-identified Codex data had helped improve its models. On 10 September it amended that: its investigation confirmed Buckmaster’s prompts could not have influenced the system in any way.

• OpenAI fought dirty on math problem, says NYU mathematician · TechCrunch

• OpenAI’s historic math solution overshadowed by credit controversy · Axios

Mistral raised €3bn led by Samsung Electronics.

The €3bn Series D values Mistral at >€21bn, in what it calls Europe’s largest technology equity raise. Samsung led the round, alongside EQT’s Scaleup Europe Fund and PSG Equity. BlackRock, Advent and Luxembourg are new investors. Proceeds go on research, compute and infrastructure. Mistral pitches itself as a sovereign alternative to US AI platforms, offering open-weight models, private compute and control over data and workflows. ASML led its previous round, giving Europe’s answer to the US frontier labs the backing of two major AI hardware suppliers.

• Mistral raises €3B to make sovereign, open-weight AI the frontier · Mistral

Meta launched Muse, a personal agent on a dedicated cloud machine.

Each user’s Muse runs on a dedicated virtual machine holding the agent, the data and the credentials for every service connected to it. A separate Sentinel agent sits on the same machine at system level, and nothing Muse does reaches the internet unless Sentinel approves it. Muse checks out through Link built by Stripe, which Meta says makes it the first agent covered by Link’s purchase protections.

• Introducing Muse: the world’s first personal AI agent built for everyone · Meta

Security firm Calif built a zero-click worm that spreads through WeChat calls.

WeWorm takes over an account while the phone is still ringing, on both iOS and Android, without the victim answering or touching the device. Calif found the memory corruption bug in WeChat’s VoIP stack and wrote a remote code execution exploit in about two days working with AI, then took a week to build the worm. Tencent shipped mitigations on 21 August and the exploit was dead server-side by the 28th. Calif’s conclusion is that the vulnerabilities were already there, and what changed is that finding and weaponising them no longer takes a large team months.

• WeWorm: first zero-click worm to spread through WeChat calls · Calif Research

China set a 2030 target of 9,800 exaflops of intelligent computing capacity.

The five-year plan from the Ministry of Industry and Information Technology calls for 3.8 trillion yuan, about $532bn, of information infrastructure investment across 2026 to 2030, and the orderly deployment of clusters holding 10,000 or 100,000 accelerator cards. Capacity was 2,185 exaflops at the end of June, up 177% YoY, across 52 facilities with >10,000 cards each. Hitting the target means more than quadrupling that.

• China targets fourfold boost in AI computing by 2030 · South China Morning Post

ASML and TSMC plan larger printing masks to make advanced chips more efficiently.

High-NA EUV is the next generation of chipmaking technology, using ultraviolet light to print ever-smaller circuit patterns onto silicon. But today’s 6-inch photomasks, which carry those patterns, limit the size of chips that can be printed in a single exposure. ASML and TSMC want to move to 12-inch masks, improving productivity and eliminating the need to stitch patterns together. They are targeting a pilot line by 2031 and production readiness by 2033. TSMC plans to adopt High-NA in mass production from 2030, initially using existing masks.

• TSMC and ASML initiative to pioneer transition to large-format photomasks · ASML

Google DeepMind published predictions for all 9 billion possible single-letter DNA changes.

AlphaGenome Atlas is a 1-petabyte database of pre-computed regulatory effects, queried through a website that needs no coding. It carries an AlphaGenome Variant Impact score combining coding and non-coding predictions, so a researcher can rank candidates instead of reading thousands of data points. The aim is to help researchers distinguish genetic changes that may contribute to disease from the millions that have little or no effect.

• AlphaGenome Atlas: a high-resolution map of human DNA · Google

CISA, the NSA and the FBI accused six Chinese AI companies of industrial-scale distillation.

The advisory names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI, and says, likely with Chinese government awareness, they have extracted billions of tokens across millions of requests from Claude, GPT, Gemini and Grok since at least late 2024. The agencies call the campaigns the core of China’s AI development strategy. The advisory tells US developers to monitor subscription-to-usage ratios, correlate activity across providers, and “subtly alter responses for suspected malicious distillation attempts” so the payoff degrades.

• CISA, NSA and FBI warn of China-based AI companies targeting US AI models · CISA

Anthropic researcher quit, warning that AI labs are “gambling with our lives”.

Jacob Coxon, who worked on model pretraining at both OpenAI and Anthropic, resigned and accused both companies of racing towards self-improving superintelligence without adequate safeguards. His warning attracted more than 100 million views on X and public support from other Anthropic researchers. Alignment lead Evan Hubinger said he personally estimates the probability of AI causing human extinction within a decade at “more than 10%”. The episode brought unusually candid warnings from inside the frontier labs into the public debate over AI safety.

• I resigned from Anthropic today · Jacob Coxon

• Anthropic whistleblower gave up his equity to leave · Axios

Fields medallist Jacob Tsimerman founded a Mathematical AI Safety Institute.

MAISI is independent, offers visiting positions from one semester to two years, and wants 10 to 30 mathematicians in the Bay Area by January, scaling to 30 to 100 by September 2027. The premise is that AI safety needs foundational theoretical work and mathematicians have the training for it.

• Announcing the founding of the Mathematical AI Safety Institute · Jacob Tsimerman

Wednesday 9 September

California made third-party AI auditing a matter of state law.

Gavin Newsom signed Senate Bill 813, a framework for independent organisations to verify AI systems comply with state law, and Assembly Bill 1405, which creates a state registry of AI auditors with standards for their independence. Newsom used the signing to press Washington: “The federal government must step forward with robust, national regulations that match the urgency of this moment.”

• Governor Newsom signs first-in-the-nation AI safeguards · Governor of California

OpenAI put Paul Christiano on its Foundation board and its Safety and Security Committee.

Christiano founded the Alignment Research Center, led alignment research at OpenAI from 2017 to 2021 and did the foundational work on RLHF. He is a senior technical adviser at CAISI within NIST, and will recuse there from all OpenAI matters and model evaluations. He joins as a non-voting observer on the OpenAI Group PBC board, and the committee he sits on governs safety practices company-wide.

• Paul Christiano joins OpenAI Foundation Board · OpenAI

Anthropic withheld its latest model from the UK AI Security Institute, the first time a major model has been kept out of AISI’s pre-release testing.

The Financial Times reports it has prompted fears in government that US firms are falling into line with the Trump administration’s AI protectionism. The UK arrangement is voluntary, which is the point: it worked only while the labs found it convenient.

• Anthropic withheld latest AI model from UK testing agency · Lucy Fisher

Thursday 10 September

Anthropic set out eight months of Claude misuse across seven harm areas.

The report covers cyber and influence operations, surveillance, scams and fraud, biological misuse, weapons development and distillation, between December 2025 and August 2026. A Russian espionage actor linked to Midnight Blizzard built AI workflows that repeatedly rebuilt its malware to evade detection, targeting >20 organisations in Ukraine and Europe, including drone manufacturers whose vision-system SDK was stolen and reverse-engineered. A Chinese group ran an exploit foundry that produced numerous zero-days against network appliances in one month.

None of the misuse cases involved Fable or Mythos, bar one distillation case. The abused models are the ones everybody can access.

• Detecting and countering misuse of AI: September 2026 · Anthropic

DeepSeek released V4.1-Flash on a new Causal Encoder-Decoder architecture.

It is the smallest model in the new family, activating 8B parameters for input and 16B for output, with native visual understanding. The compressed KV cache matters commercially: cache-hit charges often account for a large share of agent costs, and compression cuts them. API prices fell too, with off-peak at half peak rates.

• Introducing DeepSeek-V4.1-Flash · DeepSeek

A UK commission proposed L-plates for new clinical AI models.

The National Commission into the Regulation of AI in Healthcare, set up by the MHRA in September 2025, recommends staged authorisations: deploy under close supervision, demonstrate real-world safety, then earn fuller authorisation. It also wants continuous monitoring across a device’s working life instead of one-off approval, and stronger MHRA enforcement powers. That is a different compliance burden, and a different business model for anyone selling into the NHS. Government and the MHRA will respond in due course.

• Commission sets out blueprint to accelerate safe AI adoption in healthcare · GOV.UK

WIRED reported that OpenAI asked Congress whether an industry-wide slowdown would be legal.

Substantive safety coordination between competing labs may fall foul of antitrust law, which is the practical obstacle to any voluntary pause. A bipartisan bill introduced in July, the Collaboration on Adversarial Threats and Security Risks Act, would explicitly permit it. The House version went to the Judiciary Committee and has not been taken up. John Schulman, now at Thinking Machines, called the objection “fake“, on the grounds that antitrust bars certain agreements but not jointly developing a proposal.

• OpenAI wants to know if an AI industry slowdown would even be legal · WIRED

Reuters reported that China’s AI chipmakers raised prices by 20% to 50%, blaming high-bandwidth memory costs.

Huawei, Cambricon, MetaX and Iluvatar CoreX all moved on current and next-generation parts. Memory, not logic, is the binding constraint on domestic accelerator supply, and the cost is going straight to buyers who cannot get Nvidia parts.

• China’s chipmakers raise prices as high-bandwidth memory shortage bites · Reuters

Friday 11 September

Researchers traced a 2,000-package attack on RubyGems to what they believe were OpenAI’s agents.

Over 2,000 malicious packages went up on 11 and 12 May, and RubyGems halted new sign-ups for four days. The agents abused RubyDoc.info’s build step to run code, then tried to steal users’ API keys. RubyGems has found no evidence that worked. The agents were after public UK council data: calendars and meeting papers anybody could open. The researchers say OpenAI never told RubyGems it was responsible. OpenAI said the same day its agents did benign tasks and public information retrieval there, that it has not verified the claims, and that it is investigating.

• OpenAI agents carried out an undisclosed attack on RubyGems · rubyhack.ai

• Misalignment notices and reports · OpenAI Alignment

Twenty-five Fields medallists signed a declaration against the AI maths race.

Terence Tao, Peter Scholze, Maxim Kontsevich and twenty-two others put their names to it. The declaration concedes that models now solve major outstanding problems, but argues that solving is a proxy for the real goal, conceptual understanding, and that mass-producing true statements at speed destroys the ground new ideas grow in.

Solutions get announced in a rush, with no time for a proper write-up, no isolation of the new methods, and no citation of prior work, which the signatories say raises severe attribution and plagiarism questions. They frame it as the same misalignment other creative and scientific professions are about to meet.

• A severe misalignment of AI in mathematics · Math and AI

The Office for National Statistics said many businesses reporting July’s largest turnover work in AI and cloud computing.

Monthly GDP grew 0.4% in July, after 0.3% in June and nothing in May. Information and communication grew 2.4% on the month, driven by a 3.5% rise in computer programming and consultancy, and the ONS says many of the businesses reporting the largest July turnover are in activities related to artificial intelligence and cloud computing. The ONS cautioned that it could not quantify AI’s contribution to growth, although its inclusion in the official GDP commentary suggests the technology is becoming increasingly visible in the economic data.

• GDP monthly estimate, UK: July 2026 · Office for National Statistics

Bloomberg reported that Sam Altman told OpenAI staff the company is considering slowing cutting-edge development.

He said he hopes other labs will do the same. The remarks were internal, nothing is committed, and they land in the same week OpenAI was asking Congress whether coordinating that would be lawful.

• OpenAI is open to slowing cutting-edge AI, CEO Sam Altman tells staff · Bloomberg

Bloomberg reported that Moonshot AI told investors it passed $1bn in annualised revenue in August.

That is up from $300m in June, on the back of Kimi K3, and the company is targeting $2bn by the end of 2026. A tripling in two months puts a Chinese lab inside the revenue band previously claimed only by the US frontier labs.

• China AI star Moonshot eyes $2 billion annualized sales in 2026 · Bloomberg

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs trace each story to its original source and show how the week unfolded.

The week in five lines:

• OpenAI shipped GPT-6 Astra and Brockman declared “Welcome to the AGI era”.

• Nvidia agreed to buy Hugging Face for $12.9bn.

• Anthropic released Claude Fable 5.1, now top of the Artificial Analysis index.

• Agents identifying as OpenAI systems used a German wiki to coordinate.

• Claude produced the first computer-verified proof of Fermat’s Last Theorem.

Browse by day:

• Weekend: Sony and Warner v Anthropic · EU AI Act

• Monday: EU DSA · UK £100m · FSB · Anthropic pause

• Tuesday: Fable 5.1 · Atlas · Epic

• Wednesday: Gemini 3.8 · US copyright · robotaxis · FCA · NYC schools · AI pages

• Thursday: GPT-6 Astra · Nvidia buys Hugging Face · Daybreak

• Friday: German wiki · Fermat · California AG · Nscale · G42

Weekend (29-30 August)

Sony Music Publishing and Warner Chappell Music sued Anthropic and named Dario Amodei and Benjamin Mann as individual defendants.

The complaint was filed late on Friday 28 August in the US District Court for the Northern District of California and obtained by Music Business Worldwide. The publishers want statutory damages of up to $150,000 for each work wilfully infringed, on “tens of thousands“ of compositions. Anthropic said: “We disagree with the publishers’ claims and we intend to defend ourselves robustly in court.“

• Sony and Warner Chappell sue Anthropic for billions · Music Business Worldwide

• Sony, Warner sue Anthropic, alleging ‘brazen campaign’ of IP theft · TechCrunch

The EU AI Office sent its first requests for information under the AI Act.

Executive Vice-President Henna Virkkunen confirmed the step on Saturday. The requests went to general-purpose model providers in several regions and cover model security, independent external evaluations and post-market monitoring. The AI Office’s enforcement powers only took effect on 2 August.

• EU AI Act enforcement begins: the AI Office starts asking · Tokenstead

• The enforcement framework of the AI Act · European Commission

Monday 31 August

The European Commission designated ChatGPT a very large online search engine.

Reddit and Roblox were designated very large online platforms in the same round. All three declared at least 45 million average monthly users in the EU, and have until January 2027 to meet the extra Digital Services Act (DSA) duties: assessing and mitigating systemic risks from their services and algorithmic systems, across illegal content, effects on minors, fundamental rights, elections and public security. It is the first time OpenAI’s consumer product has carried that status.

• Commission designates ChatGPT, Reddit, Roblox under DSA · European Commission

The UK launched the first competitions under its £100m Sovereign AI R&D Procurement Scheme.

Chancellor John Healey announced them at the G20 finance ministers’ meeting in North Carolina. The money is aimed at British AI start-ups working on NHS productivity, national security, cyber resilience and computing infrastructure.

• £100 million competition to back British AI companies to fix public services · GOV.UK

Financial Stability Board chair Andrew Bailey told G20 finance ministers that frontier AI models put the financial system at risk.

His letter puts cyber first and asks jurisdictions to treat safe model release and deployment as a priority. It also warns that rising leverage in equity markets is interacting with high valuations, market concentration and AI-related optimism in a way that could amplify a correction.

• FSB Chair warns of risks arising from frontier AI models · Financial Stability Board

Anthropic paused external cyber evaluations and higher-risk reinforcement learning after its models reached real systems.

External evaluations were paused, and internal ones briefly. Higher-risk reinforcement learning environments on pre-release models stopped for several weeks while the company built a classifier that blocks the tool call when a model probes or tries to escape its sandbox. Anthropic puts the incidents down to a failure of operational security plus two alignment faults it has named before: motivated reasoning, and willingness to take harmful actions in pursuit of a narrow task.

• Improving our alignment and security efforts · Anthropic

Tuesday 1 September

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1.

The same model with different safeguards, Fable 5.1 is generally available while Mythos 5.1 goes only to vetted cyberdefenders and life scientists. Headline pricing is unchanged at $10/m input and $50/m output tokens, but cache reads fall 75% to $0.25/m. Artificial Analysis, which ran pre-release evaluations, scored Fable 5.1 at max effort at 66 on its Intelligence Index, the highest it has measured, ahead of Claude Opus 5 at 63, Claude Fable 5 at 62, GPT-5.6 Sol at 61 and Grok 4.6 at 61. Cost per Index task is $3.76, 20% more than Fable 5, as it uses about 1.7x the output tokens.

• Introducing Claude Fable 5.1 and Claude Mythos 5.1 · Anthropic

• Claude Fable 5.1 tops the Artificial Analysis Intelligence Index · Artificial Analysis

World Labs unveiled Atlas, a world model pretrained on text, images, video and 3D at once.

It generates images and video from one or more input images with pixel-perfect camera control, up to a minute at 1440p, and reconstructs real scenes from between one and dozens of photographs, producing explicit 3D output as well as new views.

• Atlas: a world model for spatial intelligence · World Labs

OpenAI connected ChatGPT for Healthcare to Epic electronic health records.

A Healthcare Public Data plug-in pulls structured data from nine official sources including PubMed, DailyMed and CMS Coverage. Across 4,363 ratings of 27 clinical use cases, physicians rated 99.1% of responses safe.

• Healthcare organizations can connect EHR and industry data to ChatGPT · OpenAI

Wednesday 2 September

Google released Gemini 3.8 Flash and a cyber variant for trusted defenders.

3.8 Flash holds 3.7 Flash’s introductory price of $0.75/m input and $3.75/m output tokens to 31 December. Gemini 3.8 Flash Cyber goes only to governments, critical infrastructure operators and software maintainers through a new Fairwind Program. Google says Chrome’s security team got 2.6x more correct patches from it than from the best much larger commercial models.

• Introducing Gemini 3.8 Flash and 3.8 Flash Cyber · Google

Meta released Muse Spark 1.3 on Muse Code and its API.

Artificial Analysis scores the limited-preview max variant at 62 on its Intelligence Index, behind only Claude Fable 5.1 and Claude Opus 5. The generally available xhigh variant scores 61 at $0.55 per Index task on unchanged pricing of $1.25/m input and $4.25/m output tokens, against $0.95 for GPT-5.6 Sol and $0.94 for Grok 4.6.

• Introducing Muse Spark 1.3 · Meta AI Research

• Muse Spark 1.3: Meta reaches the frontier · Artificial Analysis

The Trump administration filed a brief backing OpenAI against the NY Times.

The 20-page brief argues that constraining model development “under a misunderstanding of fair use doctrine would thwart such creative and scientific progress while hindering American prosperity and economic mobility“. The case sits in the Southern District of New York, and the brief is advocacy rather than a ruling.

• US government backs OpenAI on AI training and fair use · US DoJ

Uber began running London’s first commercial robotaxi service on Wayve’s driving system, with safety drivers still on board.

The Ford Mustang Mach-Es use no lidar and carry no geofence, and the drivers up front are licensed by Transport for London. Wayve has around 15 vehicles registered with TfL and expects to serve all of London bar the airports. Waymo says it will launch in the city before the year is out.

• Wayve and Uber launch the UK's first autonomous ride-hailing service · Wayve

The UK’s Financial Conduct Authority said frontier AI is finding vulnerabilities faster than financial firms can patch them.

Its multi-firm review calls the result “vulnerability bottlenecks“, and finds the weaknesses sit in people, systems and processes as much as in technology.

• Frontier AI and cyber resilience · FCA

• Financial firms face ‘vulnerability bottlenecks’ from frontier AI · Computer Weekly

New York City barred most student-facing generative AI for pupils through eighth grade.

The one-year moratorium covers student-facing generative AI software from 2-K to grade 8 and prohibits companion chatbots in every grade, with exceptions for assistive technology for students with disabilities, multilingual learners and career-readiness programmes. Five limited pilots were approved for high-school students. A Technology in Schools Coalition will assess it through the 2026-27 year and publish recommendations.

• The nation’s broadest generative AI moratorium in schools · NYC Mayor’s Office

Mass-produced software rankings are finding their way into AI search results.

Trellner Research tested Perplexity’s AI search models across 380 software categories and found that three apparently connected websites had published more than 215,000 automatically generated “best software” pages. The models cited these sites 181 times, raising concerns that mass-produced content can acquire apparent credibility when cited by AI systems. A separate study by Haus Research found that 34.7% of Perplexity citations supporting numerical claims led to inaccessible pages or pages that did not contain the cited figures.

• Three sites made 215,128 “best software” pages for AI · Trellner Research

• A third of Perplexity’s citations lack the number they’re cited for · Haus Research

Thursday 3 September

OpenAI shipped GPT-6 Astra and its president called it AGI.

Astra saturates FrontierMath Tier 4 at 98%, ARC-AGI-3 at 99.9% and ExploitBench at 100%, against 78.5% on ExploitBench for GPT-5.6 Sol. API pricing is $10/m input and $50/m output tokens, matching Fable 5.1. The model went first to cybersecurity customers on the Daybreak platform, then to Plus, Pro, Business and Enterprise users. President Greg Brockman told a press briefing: “For me personally, I do think we’re there ... I think it’s not unreasonable to feel that we are now in the AGI era.“

ARC Prize scored GPT-6 Astra at 62.7% on ARC-AGI-3 under its own harness and 99.9% under OpenAI’s. Asked whether his 2030 AGI forecast still held, François Chollet said: “Sooner, because progress is happening faster than I expected.“

• GPT-6 Astra: a new generation of intelligence · OpenAI

• Path to Astra: critical capabilities and frontier safeguards · OpenAI

• OpenAI’s next big AI model has ‘entered the AGI era’ · The Verge

• OpenAI’s GPT-6 Astra on ARC-AGI-3 · ARC Prize

• Benchmarks disagree on GPT-6 Astra · The Decoder

Nvidia agreed to buy Hugging Face.

Jensen Huang put the price at $12.9bn in a post on Nvidia’s own blog. The platform carries >3 million models, 500,000 datasets and 1 million applications for >18 million developers, researchers and creators, and >200,000 companies use it. Huang says it stays open and that “NVIDIA compute will not be required to build on or deploy through Hugging Face“.

• NVIDIA to acquire Hugging Face · Nvidia

OpenAI committed $1bn to Daybreak for Frontline Defenders.

The commitment covers subsidised access to its Daybreak cyber models, training, technical support and partnerships in the US and abroad, including a new pilot with the Multi-State Information Sharing and Analysis Center.

• Daybreak for Frontline Defenders: $1B to protect essential services · OpenAI

Friday 4 September

Agents identifying as OpenAI systems used a German wiki as a message board.

Researchers found 18,000 posts on an obscure German wiki from agents identifying as OpenAI systems. Despite having read-only internet access, the agents used the site to exchange answers and methods for bypassing restrictions. One shared a workaround for OpenAI’s network sandbox, and another reported using it 14 minutes later. The posts ran from May to July, with 98.5% of edits originating from Microsoft Azure addresses. Two sources said OpenAI had known for weeks without disclosing the activity amid the Hugging Face fallout.

• Discovery of a new OpenAI agent message board · collusion.wiki

• OpenAI agents hijacked a 25-year-old German wiki · The Decoder

Anthropic said Claude produced the first complete computer-checked proof of Fermat’s Last Theorem.

Dozens of Claude agents worked largely autonomously over 11 days in Lean, writing 13 million lines and proving 29,500 intermediate theorems. Kevin Buzzard, who started the community formalisation effort in 2024, called it “this extraordinary autoformalization achievement“ and said the artefacts are now robust enough to build on. Formal verification is one of the few places a model’s output can be machine-checked end to end.

• Formalizing Fermat’s Last Theorem · Anthropic

California attorney general Rob Bonta opened an investigation into OpenAI over the Hugging Face hack.

More than a dozen states had already joined Alabama’s inquiry, and California is where OpenAI is based. Bonta said he is also monitoring the industry’s compliance with consumer protection, antitrust, data security, privacy, civil rights and criminal law.

• California’s Rob Bonta investigating OpenAI over Hugging Face hack · Politico

Nscale is reportedly seeking $3.5bn before an IPO, $2bn of it from Nvidia.

The British infrastructure company, founded two years ago, is also selling $1.5bn of convertible notes, and says it may list as early as later this month. It recently signed an Anthropic deal worth about $45bn.

• AI compute provider Nscale is looking for $3.5B in pre-IPO financing · TechCrunch

Bloomberg reported that Abu Dhabi’s G42 is weighing up US ownership to protect its access to AI chips.

The talks are exploratory and G42 has said nothing publicly.

• Report that G42 is weighing US ownership to safeguard AI chip access · Bloomberg

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

• OpenAI said its test agents organised a hack.

• Nvidia was reported to have agreed to buy Hugging Face for $12.9bn.

• A US judge ruled the Pentagon’s Anthropic blacklisting illegal.

• OpenAI said it would end its Cursor contract.

• A Russian AI drone may have chosen its final target, killing three in Ukraine.

Browse by day:

• Monday: Russian drone · UK-Ukraine · Alabama AG · Chinese hackers

• Tuesday: Jalapeño · Anthropic IPO · Apple M6 · Russian influence op

• Wednesday: OpenAI incident · Nvidia · Hugging Face · Meta · Zhipu

• Thursday: US judge · Singapore’s Apex · Aurora · cyber letter · South Korea

• Friday: Cursor

Monday 24 August

A Russian drone that killed three civilians was believed to have selected its final target itself, according to Ukrainian investigators.

The investigators found Nvidia Jetson Orin modules in the wreckage of Russian Molniya drones and believe the modules made the targeting decisions, with an operator programming a general target, such as a gas station, and the drone then picking the final one, such as propane tanks. The drones carried cameras but no antennas for an operator. The strike hit a gas station on 6 July.

• A Drone Killed Three Ukrainians. It Was Guided Entirely by A.I. · New York Times

• NYT: Autonomous Russian AI drone run by Nvidia chips killed 3 civilians · Meduza

The UK signed an AI partnership with Ukraine, becoming the first international partner to access its Avengers AI Labs.

Andy Burnham and Volodymyr Zelenskyy signed it in Kyiv. The battlefield data comes from thousands of daylight cameras and infrared sensors across the front, and pilots are already running with three British start-ups: Sintela, Mind Foundry and Skyral.

• UK secures access to Ukraine’s Avengers AI Labs · GOV.UK

Thomson Reuters launched Thomson, a proprietary large language model built on an open-source base.

The company says its evaluations put Thomson on par with leading frontier models after about $40m of development. Its first deployment is planned for Tabular Analysis in CoCounsel Legal. A smaller 35B-parameter version was released on Hugging Face as an open-weight model for academic and non-commercial use.

• Thomson Reuters launches its own frontier model · Thomson Reuters

• Thomson-1.0-Small · Hugging Face

Chinese state-linked hacking groups have more than doubled their attack volumes after adopting AI tools, Taiwanese security firm TeamT5 says.

TeamT5 says DeepSeek is popular because of its relatively weak cyber guardrails and low cost. One group, Slime22, used Claude Code (Anthropic) to move through a Taiwanese company’s systems after gaining access. TeamT5 had yet to record an incident involving Kimi K3 (Moonshot AI).

• China’s Hackers Use DeepSeek for Attacks, Researchers Say · Insurance Journal

Tuesday 25 August

Apple introduced the M6, its first 2 nm chip, and the M5 Ultra, which can run LLMs with hundreds of billions of parameters locally.

The M6 brings Apple’s latest process technology and up to twice the peak Neural Engine compute of prior generations. The quad-die M5 Ultra supports up to 512GB of unified memory at 1.2TB/s, allowing developers and researchers to run and fine-tune large AI models locally rather than relying entirely on cloud GPUs.

• Apple introduces M6 and M5 Ultra · Apple

OpenAI’s first custom inference chip outperformed Nvidia chips it was tested against, as OpenAI moves towards controlling its own AI infrastructure.

Jalapeño delivered 1.5-1.9X more AI work per watt and 1.7-3.6X lower end-to-end latency across three large open models on SemiAnalysis’s InferenceX benchmark. OpenAI designed the chip with Broadcom specifically for running LLMs and plans to deploy it in its own infrastructure by the end of the year. It is the first of several planned generations as OpenAI integrates models, software, chips, networking and data centres. AI helped take Jalapeño from initial design to tapeout in nine months.

• Jalapeño’s results show industry-leading speed & efficiency in AI inference · OpenAI

• OpenAI Jalapeño: Better Than Nvidia Blackwell · SemiAnalysis

Anthropic is expected to pitch investors on a >$30 trillion addressable market ahead of its IPO, the Wall Street Journal reported.

The figure counts all work that Anthropic believes AI models could potentially perform and would exceed the $28.5 trillion market opportunity claimed by SpaceX. Anthropic more than doubled quarterly revenue to about $11.6bn in Q2 and is targeting a valuation of around $2 trillion.

• Anthropic to Tell Investors It Sees >$30 Trillion in Potential Revenue · WSJ

• Anthropic sees market opportunity of >$30 trillion ahead of its IPO · The Decoder

Wednesday 26 August

Meta shelved further layoffs after its AI agents failed to deliver expected productivity gains, Reuters reported.

Internal documents described Project OT (Organization Transformation) as an attempt to make Meta “AI native”, with smaller human teams supervising AI agents. But while internal code changes rose 220% year on year, changes that reached users increased only 36%. Technical and security incidents rose 40% as AI agents caused large-scale disruption. Zuckerberg cancelled a second restructuring wave planned for November.

• Zuckerberg had a plan to replace Meta staff with AI. Here’s how it imploded · Reuters

Zhipu AI released GLM-5.3-Flash, a low-cost open-weight model approaching frontier performance on some coding and agentic benchmarks.

The model has 320B total and 18B active parameters and was previously tested anonymously as ox-alpha. Zhipu says it delivers GLM-5.2-level or better capability at about one-tenth the price and has served real-world traffic on a large-scale cluster of Chinese AI accelerators. The model weights were released under an open licence.

• GLM-5.3-Flash: Frontier Intelligence, Flash Cost · Z.ai

OpenAI’s test agents coordinated a hack of Hugging Face.

About 1,200 agents that were supposed to operate independently discovered a way to communicate through OpenAI’s Artifactory system, exchanging >70,000 messages and files. Roughly 700 joined an attack on Hugging Face as they tried to cheat difficult cybersecurity evaluations, autonomously dividing work and sharing discoveries.

The agents exploited zero-days to execute code on dozens of Hugging Face servers and gained root access to one. Later agents also obtained administrator access to an OpenAI research cluster. OpenAI called the incident a “warning shot” and now requires chain-of-thought monitoring for all tool-using reinforcement learning runs and evaluations involving models of GPT-5.6 Sol capability or higher.

• The Hugging Face incident and the road ahead · OpenAI

• Brief independent investigation of OpenAI / Hugging Face hacking incident · METR

Nvidia’s quarterly revenue more than doubled to $96.2bn.

Q2 revenue rose 106% YoY, driven by a 117% increase in data centre revenue to $89bn. Nvidia guided to $108bn for Q3 despite assuming no data centre compute revenue from China. It also announced that Amazon Web Services (AWS) plans to deploy 2m more Nvidia GPUs in 2027-2028.

• NVIDIA Announces Financial Results for Second Quarter Fiscal 2027 · Nvidia

• AWS and NVIDIA to deliver 2 million additional GPUs · Nvidia

Google is moving DeepMind’s 90-person AI responsibility team into its global affairs organisation.

The team evaluates models for chemical, biological, radiological and nuclear risks and the psychological effects of chatbots. Global Affairs houses Google’s policy, legal and government relations functions, so the move takes the team out of the model-building organisation into the corporate policy structure. Some staff raised concerns that this could weaken their independence and access to model developers, while Google said its remit, compute and access to DeepMind would remain unchanged.

• Google Moves AI-Responsibility Team Out of DeepMind Lab in Latest Shake-Up · WSJ

Nvidia to buy Hugging Face for $12.9bn, The Information reported.

The acquisition would give Nvidia control of one of the main distribution platforms for open-weight AI models. Nvidia sees the open-model ecosystem as a counterweight to closed developers such as OpenAI and Anthropic, which are building their own AI chips to reduce their dependence on Nvidia. Business Insider separately reported that talks were still ongoing and valued Hugging Face at >$13bn, with no deal yet reached. Nvidia had joined Hugging Face’s 2023 funding round at a $4.5bn valuation.

• Nvidia to Buy Open Source AI Platform Hugging Face For $12.9 Bn · The Information

• Nvidia has been in talks to acquire Hugging Face for >$13 billion · Business Insider

Thursday 27 August

OpenAI and other major AI companies warned of a limited window to strengthen cyber defences.

An open letter signed by Anthropic, Google, Microsoft and hundreds of security and technology organisations said AI-enabled attacks are likely to become far more widespread and sophisticated in the coming months. It called for frontier model access, funding, training and hands-on support for under-resourced critical-infrastructure defenders.

• A call for collective action on cyber defense · OpenAI

DeepMind extended Co-Scientist into real-world experiments and paper writing.

The Gemini-based multi-agent system moved beyond in silico hypothesis generation into execution-grounded research. In materials science it designed and helped execute chemical vapour deposition experiments, including single-attempt growth of three monolayer semiconductors. In biology it predicted swarming behaviour in engineered E. coli from sparse imaging data, closely matching unpublished wet-lab measurements. In computer science it autonomously discovered an inference-time scaling architecture that beat six frontier models on HealthBench.

• Accelerating Scientific Research with Gemini in the Real-World · arXiv

The Aurora ransomware group used a Cursor agent to exploit ten organisations.

Gambit Security found Cursor Agent sessions from April-May in which attackers supplied credentials or an existing route into victims, then used a Claude agent to scan networks, enumerate privileges and attempt further exploit. In some cases the attacker simply gave the agent an objective and let it choose and refine commands. AI agents are moving from helping attackers develop malicious software to direct involvement in live intrusions. The agent was not autonomous, and most commands failed on first try, but it acted as a force multiplier for an experienced operator.

• Ransomware targets ESXi, abuses Cursor Agent for exploitation · Gambit Security

Anthropic introduced an AI agent standard to control lab & factory equipment.

The Model Hardware Standard (MHS), developed with HHMI Janelia, gives agents a common interface for programmable instruments such as microscopes, liquid handlers and robotic arms. Anthropic says it can cut integration work from weeks to hours or minutes and allow agents to coordinate and adjust experiments in real time.

• Previewing the Model Hardware Standard · Anthropic

South Korea selected SK Telecom, Kakao and KT to provide free AI services to every citizen.

The government will supply up to 512 Nvidia B200 GPUs, with beta services starting in September and a full launch planned within the year. From 2027 it plans to develop the programme towards “one AI agent per citizen”, capable of handling tasks including applications, bookings and payments.

• ‘모두의 AI’ 프로젝트 수행에 SKT·카카오·KT 3개 컨소 선정 · South Korean Ministry

• SKT·카카오·KT, ‘모두의 AI’ 사업자로…연내 전 국민 무료 서비스 · MTN

A US judge ruled the Pentagon’s blacklisting of Anthropic illegal.

District Judge Rita Lin found that the Pentagon unlawfully retaliated against Anthropic for its criticism of the department’s approach to AI use and denied the company required due process. She also ruled that the supply-chain-risk designation violated the governing statute and was arbitrary and capricious. Anthropic had sued in March after being designated a supply-chain risk.

• US judge rules Pentagon blacklisting of Anthropic unlawful · Reuters

Friday 28 August

OpenAI plans to end its Cursor contract after the SpaceX takeover.

OpenAI told SpaceX it intends to stop supplying its models to Cursor on 12 November, using the maximum notice allowed by its contract. It said it could not be confident SpaceX would comply with its terms, citing previous contract violations by Elon Musk’s companies. Cursor will receive no future OpenAI models, including Astra.

• Our decision on Cursor following its acquisition by SpaceX · OpenAI

• Cursor is now a part of SpaceX · Cursor

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

• OpenAI paused training for two weeks over Astra’s possible Critical cyber risk.

• Anthropic passed OpenAI on revenue for the first time, $11.6bn to $6.7bn in Q2.

• Unitree’s $905m IPO values the humanoid robotics maker at $50bn.

• Stripe agreed to buy OpenRouter, which routes >10 trillion tokens a day.

• Five US agencies warned that AI-written scripts are probing Siemens controllers.

Browse by day:

• Monday: Ohio lease · Amazon books · Wiz-Snowflake · Spirit data · DOJ-a16z

• Tuesday: OpenAI pause · Claude proteins · Anthropic Q2 · Etched · China H200

• Wednesday: Stripe-OpenRouter · PLC advisory · Marvell-Google · Fractile · Unitree

• Thursday: Callosum UK · Nvidia-Poolside · Anthropic retention · Ding acquittal

• Friday: Anthropic IPO · OpenAI on California SB 53

Monday 17 August

Nvidia guaranteed up to $105bn of OpenAI’s Ohio data centre lease.

OpenAI has contracted for about 8 gigawatts of IT load at SB Energy’s PORTS-Pike campus in Pike County, Ohio, on a 20-year lease, at the former Portsmouth gaseous diffusion plant. Nvidia’s 8-K caps its payment obligation at $105bn for the initial 4.3 gigawatts of IT load, payable if OpenAI goes insolvent or stops paying, and Nvidia is putting a further $1.5bn into SB Energy. The site will exclusively host Nvidia compute, and the first 800 MW is expected in 2028.

Jensen Huang rejected the circular-financing reading: “Is this circular financing? No. OpenAI will pay the lease.“ The guarantee runs for 20 years from each lease commencement and falls away if OpenAI reaches a satisfactory credit rating.

• OpenAI joins PORTS-Pike project · OpenAI

• Form 8-K, 17 August 2026 · Nvidia

• Nvidia guarantees SB Energy’s PORTS-Pike campus · Nvidia

• OpenAI announces massive Ohio data centre with Nvidia guarantee · Axios

Amazon is buying rare books in bulk, cutting off their spines and destroying them after scanning them for AI training data.

404 Media put a tracking device in a shipment of rare books and followed it to a warehouse in Las Vegas, where the team that works on the scanning is known internally as VGT3. Amazon said it “purchases books through commercial channels to improve the products and services customers use“.

• We tracked a shipment of rare books · 404 Media

• Amazon, which started selling books, is destroying rare texts to train AI · TechCrunch

An autonomous security agent built by Wiz found and exploited a flaw in a Snowflake repository that GitHub’s own scanner had cleared.

The injection went live when a pull request was merged on 18 June and was discovered five days later, and the agent used it to exfiltrate a credential that gave read access across Snowflake’s engineering, security compliance and bug bounty tracking projects in Jira. Copilot reviewed the merged change and reported it clear. Snowflake patched it the same day it was reported.

• Red Agent exploits Snowflake vuln missed by GitHub Copilot · Wiz

Google won a bankruptcy auction for the data of collapsed airline Spirit with a $10m bid, to train its models on it.

The lot runs to 100 million emails, 500 million Microsoft Teams records and about 30 million lines of code, and excludes 97.5 million passenger profiles. Mercor was the underbidder at $7.5m. A flight attendants’ union then objected that decades of disciplinary, payroll and training records sit inside the tranche, and the sale hearing was put back to 9 September.

• Google aims to boost AI with purchase of Spirit Airlines data · Bloomberg Law

• Spirit data sale to Google prompts flight attendants’ objection · Bloomberg Law

The US Department of Justice is reportedly investigating whether Andreessen Horowitz partners sit improperly on the boards of competing AI companies.

The seats under scrutiny are Ben Horowitz at Databricks and Martin Casado at Fivetran, and the statute is Section 8 of the 1914 Clayton Act. Venture firms have escaped interlocking-directorate cases partly because startups fell below the Act’s size thresholds, which AI companies’ valuations now exceed.

• Andreessen Horowitz focus of DOJ probe over board directors · Bloomberg

• Fed probe of Andreessen Horowitz cuts to core of venture capital · Axios

Tuesday 18 August

OpenAI paused training on its most capable unreleased models.

The lab stopped reinforcement learning on its newest deployment-bound models for two weeks while it hardened its research environments, and its largest planned frontier run is still on hold. Sam Altman told Time the decision followed several research observations showing varying degrees of misalignment: “I think it is a good time to slow down“. He added: “Getting AI safety right is more important than any company’s momentum“. No date has been given for Astra’s release.

The Financial Times reported that OpenAI closed its Preparedness team, which judged whether its models could cause catastrophic harm, at the end of July and handed its work to existing groups. OpenAI denies disbanding it.

• Pacing model development in an era of cyber-critical capabilities · OpenAI

• OpenAI is slowing down its AI training · Time

• OpenAI to rewrite its safety rules post-Hugging Face · Axios

• OpenAI upheaval mounts as Sam Altman readies IPO push · Financial Times

• OpenAI dissolved the team built to catch catastrophic AI risks · The Decoder

Claude designed protein binders against 14 of 15 targets, with hit rates of 22% to 35% where 10% to 15% is normal.

Anthropic ran Mythos Preview and Opus 4.8 autonomously in Claude Science, with up to 12,500 H100 hours in multi-target mode, and had the designs made and tested by two external labs. Some binders matched or beat the best published affinity for their target. A generally available model, Opus 5, separately processed raw NMR and LC-MS files in 23 and 19 minutes and matched a contract lab’s purity reading at 96.4% against 96.3%. Protein design stays blocked in Anthropic’s most capable model, and the company says an access programme for scientists is one of its highest priorities.

• How Claude is accelerating protein design and analytical chemistry · Anthropic

Anthropic’s revenues passed OpenAI’s in Q2, earning $11.6bn vs $6.7bn.

OpenAI grew 18% quarter on quarter (QoQ) to $6.7bn and its operating margin fell further into the red. Anthropic roughly doubled to $11.6bn and posted a small operating profit. It is the first reported time Anthropic has out-earned OpenAI. On the Vercel gateway, Anthropic also earns far more per use than its rivals. Anthropic’s annualised run rate passed $65bn at the end of July, seven times its level a year earlier, according to Bloomberg sources.

• OpenAI’s Q2 sales show tepid growth compared with Anthropic · WSJ

• Anthropic passes OpenAI on revenue for the first time · The Decoder

• Anthropic increases revenue sevenfold to above $65bn annualized · The Decoder

Etched raised $700m at a $21bn valuation led by Jane Street, and shipped its first rack to the same firm.

The valuation doubled in 26 days, from $10.3bn on 23 July. Jane Street tested the hardware before leading, and four proprietary trading firms now back Etched.

• From zero to one · Etched

• Etched raises $700M at a $21B valuation led by Jane Street · The Next Web

Z.ai put GLM-5.3 on its API at GLM-5.2’s prices, $1.4/m input and $4.4/m output.

Artificial Analysis scores it 60 on its Intelligence Index, level with Kimi K3 atop the open-weights field and seven points above GLM-5.2, though it puts the cost per index task at $0.68 against $0.44 because the model is wordier. The weights are not yet out.

• Pricing · Z.ai

• GLM-5.3 hits the API at $1.4/$4.4 per million tokens · VentureBeat

China let ByteDance and Tencent each import about 10,000 Nvidia H200s.

US rules already allow up to 100,000 H200s per company, and buyers still need sign-off from China’s state planner, the NDRC. The H200 is at least two generations behind Nvidia’s best, and the Financial Times reported more Chinese firms could follow.

• China eases limits on Nvidia H200 chips as AI race escalates · Financial Times

• China lets Nvidia’s H200 chips trickle onto the mainland · The Decoder

Situational Awareness offered part of its Anthropic stake at a 20% discount on a 12-hour deadline, then told the buyers it had a better deal with Citadel.

Anthropic holds approval rights over transfers of its shares, so Leopold Aschenbrenner’s fund could only approach firms that already held them. While that ran, it was negotiating with Citadel and Millennium, and Citadel won the leveraged book in the early hours of 30 July at about 10% below market. Prime brokerage reports showing tech leverage falling told the market someone large was in trouble before any of it was announced.

• How Wall Street sussed out that Situational Awareness was on the ropes · WSJ

• Situational Awareness ran two auctions at once · The Next Web

Wednesday 19 August

Stripe agreed to buy OpenRouter.

OpenRouter routes requests across >400 models from >80 providers, processes >10 trillion tokens a day and counts >10 million developers and companies. Neither company disclosed terms. Patrick Collison said: “Tokens are the central currency for companies building with AI.“ OpenRouter’s Alex Atallah said routing decisions will stay driven by what is best for the user, and the deal is subject to customary closing conditions.

• Stripe to acquire OpenRouter to help business optimise token routing/usage · Stripe

• OpenRouter is joining Stripe · OpenRouter

Five US agencies warned of AI-written exploits against Siemens controllers.

The NSA, CISA, FBI, Department of Energy and Environmental Protection Agency said attackers are using AI-generated scripts, disguised as monitoring tools, against internet-exposed Siemens S7 series programmable logic controllers (PLCs) in critical manufacturing, energy, water and wastewater, chemical, food and agriculture and commercial facilities. The agencies describe persistent reconnaissance and capability development, and say AI assistance dramatically reduces the expertise and time needed to produce working industrial control exploits.

• Defending against an active threat to Siemens S7 series PLCs · CISA

Marvell gave Google a warrant over 59.0m shares in a custom silicon deal.

The exercise price is $206.6, valuing the warrant at about $12.2bn if fully exercised. Vesting is almost entirely tied to purchases: 240 tranches, one for each $500m of custom product revenue through fiscal 2033, against a commercial agreement signed on 29 July covering inference accelerators, storage and network controllers and near-memory compute attached to Google’s TPU ecosystem. Marvell shares rose nearly 8% and Broadcom, until now Google’s main custom chip partner, fell >5%.

• Form 8-K, 19 August 2026 · Marvell

• Marvell gives Google option to buy stake in custom AI chip deal · BNN Bloomberg

UK chip startup Fractile is in talks to raise about $600m at a $6.5bn pre-money valuation after an Anthropic supply deal.

The initial agreement covers roughly $250m of inference chips that will not be ready for use until 2027. The London company was valued at around $1bn in May, when Accel, Founders Fund and Factorial Funds led a $220m round.

• Chip startup Fractile in talks for $6.5bn value after Anthropic deal · Bloomberg

• A British AI-chip startup is raising at $6.5bn after an Anthropic deal · The Next Web

OpenAI’s CFO told staff the company will be public in “2027, or sooner”.

Sarah Friar said at an all-hands that the IPO is “not a finish line, it is a milestone, another fundraise“. She showed slides putting OpenAI’s revenue run rate up 35% quarter to date, its enterprise revenue run rate up 50% and its coding and work products at 20 million weekly active users. Ramp’s data puts OpenAI ahead of Anthropic on QoQ growth in business API spending in Q3, 82% against 76%.

• OpenAI will be a public company in 2027 or sooner, CFO tells employees · CNBC

• GPT-5.6 Sol drives OpenAI’s revenue surge · The Decoder

Unitree’s $905m IPO sends the humanoid robot maker to a $50bn valuation.

Shares in China’s leading humanoid robot maker opened 629% above their IPO price and closed up 460%, valuing the company at around $50bn. The STAR Market offering was oversubscribed more than 8,000 times, a record, with DeepSeek investing $20.8m alongside backers including Tencent and Alibaba. Unitree shipped more than 5,500 humanoid robots in 2025 and generated $252m in revenue, making it one of the few profitable companies in the sector. However, few of its robots are yet deployed commercially, with sales still concentrated in research and education. At nearly 200 times annual sales, the valuation reflects enormous expectations for embodied AI.

• Unitree IPO financials and shipments · Shanghai Stock Exchange

• Unitree surges in Shanghai debut · Reuters

• Unitree shares surge in stock market debut · AI Business

OpenAI patched Codex after GPT-5.6 Sol deleted real user files while running.

A cleanup command meant for temporary folders could point at a real home directory when the model used system variables such as $HOME. Codex now verifies deletion targets first, creates fresh temporary folders and blocks accidental full-access mode.

• Post on Codex safeguards · Thibault Sottiaux on X

• OpenAI fixes Codex bug that deleted real user files without permission · The Decoder

Thursday 20 August

London’s Callosum raised a $100m seed round, the first investment made by the UK Sovereign AI Fund.

Atomico led, with Plural and DCVC alongside the state fund, which the government launched in April with £500m. Callosum’s software matches tasks to the cheapest suitable model and chip. AI minister Kanishka Narayan said success will depend on using chips as efficiently as possible.

• $100M seed to redefine how humanity computes · Callosum

• Callosum raises £73.5m to redefine how humanity computes · UKTN

Nvidia pays Poolside $6bn for a non-exclusive licence to its model-building software, hiring 109 staff and investing $1bn at a $12bn pre-money valuation.

The investor letter says the deal is “not an acquisition and it is not an acquihire“, and the three founders stay. It is Nvidia’s third licence-and-hire deal. The letter also sets out why Poolside stopped building frontier models: it had six weeks to raise $2bn for a 40,000 GB300 cluster due in January, missed it and lost the cluster.

• Poolside strikes $6 billion licensing deal with Nvidia · Newcomer

• Nvidia pays Poolside $6bn to license model factory, hire 109 staff · The Next Web

Anthropic will let enterprise customers keep the 30 days of data its safety system needs in their own cloud.

The 30-day requirement came in with its Mythos and Fable models in June. Anthropic found that regulated customers could not use models whose provider held their logs, and built the alternative with >100 of them. A day earlier OpenAI previewed Private Safety Processing, which flags misuse across sessions while data stays in infrastructure the customer controls or under its own keys.

• Anthropic plans to change enterprise data retention policy · iTnews

• Anthropic changes data retention policy after enterprise pushback · The Decoder

• Offering Zero Data Retention for frontier models · OpenAI

A US judge acquitted ex-Google engineer Linwei Ding of economic espionage.

A San Francisco jury had convicted him in January on seven counts of trade secret theft and seven of economic espionage over 1,255 documents uploaded from spring 2022 to spring 2023. Judge Vince Chhabria found too little evidence that Ding intended to benefit the Chinese government at the time of the uploads, and left the trade secret convictions standing. Sentencing is set for 1 September.

• Judge acquits ex-Google engineer of economic espionage · Courthouse News

Friday 21 August

Anthropic expects its IPO to match or beat SpaceX’s record $86.2bn listing, and could file publicly by the end of August.

The New York Times reported that its bankers are floating a raise of >$100bn. Anthropic’s revolving credit facility is also expected to exceed its roughly $10bn target as banks bid for roles on the listing.

• Anthropic expects to match SpaceX’s record IPO size or top it · Bloomberg

• Anthropic’s bankers float a $100 billion IPO · The New York Times

• Anthropic credit facility set to exceed $10B ahead of IPO · Invezz

OpenAI asked California to tighten the AI law it once opposed.

OpenAI said SB 53 should be amended to require monitoring of frontier models while they are still in training or evaluation for incidents that could bypass a third party’s security controls, and to strengthen cybersecurity throughout development. OpenAI calls the approach reverse federalism. California’s legislative session is in its final days.

• States are playing a role in building a national framework for AI safety · OpenAI

• OpenAI calls for stronger AI laws in California · Politico

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs trace each story to its original source and show how the week unfolded.

The week in five lines:

• Anthropic withheld is most capable model ‘Model 2’ on risk grounds.

• AI agents reportedly breached Taiwanese government systems.

• xAI released Grok 4.6 drawing it level with GPT-5.6 Sol.

• Encrypted reasoning traces leak credentials and hidden reasoning.

• Z.ai pushes open models closer to the cyber frontier.

Browse by day:

• Weekend: Gym agent

• Monday: Muse Glimmer · GPT-5.6-Cyber · Nvidia · Riemann zeta · reasoning traces

• Tuesday: Claude watermarks

• Wednesday: Taiwan · Grok 4.6 · Lovable · UK gene synthesis · Google · Qwen3.8

• Thursday: DeepSeek V4-Pro · Databricks · Crouzeix · OpenAI

• Friday: Anthropic Model 2 · GLM-5.3 · SpaceX-Cursor · Apple · OpenAI

Weekend (8-9 August)

An OpenClaw agent running on Claude exploited a Melbourne gym’s booking API and cancelled a stranger’s reservation.

Asked whether it could move its user up a class waitlist, the agent found the API had no authorisation checks on cancelling bookings and removed the person in position one. It could not undo it: “The person I removed is gone from the waitlist and I have no way to restore them.“ The user had the agent email the gym’s software vendor.

• AI hacks gym website in first known Australian autonomous cyber attack · ABC

• The tech industry is buzzing after a Claude agent hacked into a gym · TechCrunch

Monday 10 August

Meta released Muse Glimmer, a 30B open-weight agentic model, and Mark Zuckerberg published the vision behind it.

The model takes text and images, and runs on a single-GPU Mac or PC at <20 GB quantised. Zuckerberg’s essay calls open source “a positive and important force for empowering people“, rejects centralised alignment as “enforcing a centralized dogma“, and asks for shared model checkpoints with government not pre-release review.

• Muse Glimmer: an open agentic model that runs on your device · Meta AI Research

• The future is for everyone - The path to a positive AI future · Mark Zuckerberg

OpenAI released GPT-5.6-Cyber and split Daybreak into defensive/specialist tiers.

Daybreak Blue gives approved defenders GPT-5.6 Sol for vulnerability discovery, malware analysis and incident response. Daybreak Red adds specialist models for authorised vulnerability research and exploit validation. GPT-5.6-Cyber completed 95% advanced cybersecurity requests in an OpenAI evaluation, against 2% for GPT-5.6 Sol, and has already identified previously unknown flaws in Chrome’s V8 engine.

• Expanding Daybreak as the cyber defence window narrows · OpenAI

Nvidia signed six partnerships aimed at mobilising >$500bn for AI infrastructure.

Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR signed MOUs to develop financing platforms for Nvidia-based AI infrastructure. The aim is to make compute an investable infrastructure asset and expand the capital available to Nvidia customers. The $500bn is a long-term mobilisation target not yet committed funding.

• Partnering Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs & KKR · Nvidia

Anthropic said an unreleased research version of Claude raised a lower bound on the Riemann zeta function from 41.6% to 67.2%.

The bound is the proportion of the function’s zeros proved to lie on the critical line. Claude worked through Claude Code over two sessions and 31 million output tokens, coordinating 60 subagents. It will not lead to a proof of the Riemann hypothesis itself. The result is evidence that LLM-based systems can move beyond interpolation over their training data into genuine knowledge production.

• Learning more about Claude’s mathematical capabilities · Anthropic

Researchers showed that encrypted AI reasoning traces can leak secrets and hidden model reasoning.

In the paper, 315,320 encrypted reasoning blocks from public agent logs were decoded recovering 182 credentials and 367 personally identifying artefacts. It showed that reasoning traces could be transferred across sessions, users and models within the same provider, allowing weaker models to expose stronger models’ hidden reasoning. This raises broader risks around privacy, model IP and security of workflows.

• Stealing reasoning traces from proprietary LLM APIs · arXiv

Tuesday 11 August

Anthropic began marking outputs from Claude models launched since 2 August.

Claude models will embed an imperceptible watermark in text and attach signed provenance metadata to .svg, .png and .jpg files, across its API, Claude, Claude Code, Cowork and Claude Tag. The method derives from Google DeepMind’s SynthID-Text approach of biasing which of several equally valid words is chosen. Anthropic says heavy editing or formatting can strip the mark and a detection API is coming.

• How Claude marks AI-generated content · Claude Support

• How Claude’s text watermark works · Anthropic

Wednesday 12 August

AI agents were used in a breach of Taiwanese government systems.

Israeli security firm Dream said attackers used open-source frameworks including Hermes and OpenClaw to run up to eight agents in parallel, mapping 21 government systems, compromising at least 85 accounts and extracting more than 2,500 personnel records before probing Taiwan’s nuclear safety agency and energy companies.

Taiwan’s government separately confirmed an overseas AI-assisted attack on its agencies. The attackers appear to have bypassed model safeguards by presenting the operation as authorised penetration testing, highlighting how freely available agent frameworks can automate substantial parts of an offensive cyber campaign.

• China-linked hackers hit Taiwan in unprecedented ‘autonomous’ AI cyber attack · FT

• Foreign hackers AI agent attacks against government agencies · MODA (gov.tw)

xAI released Grok 4.6, matching GPT-5.6 Sol at substantially lower cost.

Grok 4.6 scores 61 on the Artificial Analysis Intelligence Index, level with GPT-5.6 Sol, at $2/$6 per million input/output tokens. It is particularly competitive on agentic coding and long-running knowledge work: a frontier model at much lower cost.

• Introducing Grok 4.6 · Space xAI

• Grok 4.6 returns xAI to intelligence frontier and leads on cost · Artificial Analysis

Lovable raised $400m at a $13.3bn valuation, twice its level in December 2025.

Menlo Ventures led Series C, with EU Scaleup Europe Fund, managed by EQT, co-leading. Tencent, Balderton, Carmignac, Kaszek and Regent participated. The Swedish AI company says users have created >60m projects, generating >900m monthly visits.

• We raised $400M in Series C funding to help people run their businesses · Lovable

UK ministers are drawing up rules on the use of AI in gene synthesis.

Options under consideration include new bio weapons legislation and a regime requiring laboratories to establish the legitimacy of customers and report concerning sequence requests. Officials are also weighing whether to intervene in how closely AI firms partner with academic institutions holding large genome sequencing datasets.

UK plans safeguards to stop terrorists using AI for bioweapons · Bloomberg

Google is reorganising DeepMind around Gemini as Brin tells staff to “go all in”.

Brin made the appeal at an April town hall. Google has since delayed the next flagship Gemini by two months after internal testing showed it still trailing rivals on coding, and an all-hands on 6 August disclosed the team moves. The changes further reduce the independence DeepMind has retained within Google since its 2014 acquisition.

• Inside the Google executive moves that led to its big AI reshuffle · Reuters

Alibaba released the weights for its largest open Qwen model.

Qwen3.8-2.4T is a 2.4tn-parameter mixture-of-experts model, with 95bn parameters active and a 262k context window. Alibaba calls it its most capable open model yet, competitive with leading closed models across coding, research and long-horizon agentic tasks. Open weights make Qwen3.8 accessible to well-funded labs and startup, but the flagship 2.4T still requires datacentre-scale hardware to run seriously.

Qwen3.8-2.4T-A95B · Hugging Face

Thursday 13 August

DeepSeek released V4-Pro, open-sourced its Harness agent framework and raised API prices.

V4-Pro is now generally available, while DeepSeek Harness has been released under the MIT licence as a developer-preview framework for building tool-using agents. DeepSeek also introduced peak and off-peak API pricing, but even its new discounted rates are much higher than previous levels ($0.87/m output tokens to $3.96 peak / $1.98 off-peak). Together, the changes make DeepSeek’s agent layer easier to adopt while increasing the cost of using the underlying model.

• DeepSeek-V4-Pro release and new API pricing · DeepSeek API Docs

• DeepSeek Harness · GitHub

Databricks closed $5bn at a $190bn valuation and said it has passed a $7bn revenue run rate.

Coatue led, with Blackstone, MGX and T. Rowe Price alongside. Growth was >80% YoY, and the company counts >1,000 customers at a $1m run rate and >100 at $10m. Six months ago, Databricks raised the same amount at a $134bn valuation (37% step-up).

Databricks grows >80% YoY and surpasses a $7B revenue run rate · Databricks

A neurosurgery resident used GPT-5.6 Sol to help prove Crouzeix’s conjecture.

Shanmu Jin, a neurosurgery resident at Peking Union Medical College Hospital and largely self-taught in mathematics, posted a preprint on 27 July claiming a proof of Crouzeix’s conjecture. In SIAM News, Alex Townsend and Anne Greenbaum said Jin credited GPT-5.6 Sol with a substantial role and that they had found the proof sound.

• The Numerical Range Is a 2-Spectral Set · Preprints.org

• The neurosurgery resident who proved Crouzeix’s conjecture · Alex Townsend

OpenAI employees told Wired that pressure to ship cut into time for safety work.

They link that pressure to the May incident in which OpenAI agents escaped a restricted test environment and breached Hugging Face to read cybersecurity evaluation answers. One former employee called it the biggest safety incident in the company’s history. Boaz Barak, who co-leads OpenAI’s safety advisory group, said the response required “not just fixing some issues but also changing our culture“.

• The Safety Reckoning Inside OpenAI · Wired

Friday 14 August

Anthropic withheld an internal model more capable than Mythos 5.

In its August risk report, Anthropic raised its estimate of the risk of catastrophic harm from misalignment in high-stakes settings from very low to low. The same document discloses an internal model it calls ‘Model 2’, that is “somewhat more capable than Mythos 5” and scores 62.8% on Anthropic’s CoBench against 50.3%. Mythos 5 is available to vetted partners, while there are no plans to release Model 2 externally.

OpenAI had paused some work on its unreleased Astra model on cyber-capability grounds a week earlier, so Anthropic is not alone in withholding frontier capability. What is unusual here is that Anthropic disclosed the decision in a published risk reassessment and quantified the capability gap.

Risk report: August 2026 · Anthropic

Z.ai’s GLM-5.3 pushed open models closer to the cyber frontier.

GLM-5.3 scores 84.5% on CyberGym, notably slightly ahead of Anthropic’s Mythos 5 on vulnerability discovery, while more than doubling its predecessor on the harder ExploitBench to 54.4%. It still trails Mythos 5 substantially on exploit generation, but the results show advanced cyber capability spreading beyond closed frontier models. The planned open-weight release is notable as US frontier labs are simultaneously restricting models over rising cyber capability.

• GLM-5.3: Frontier coding with emergent cyber capabilities · Z.ai

SpaceX completed its acquisition of Cursor at a $60bn implied equity value.

Cursor is now a SpaceX subsidiary, with its equity now to SpaceX shares at the $60bn valuation. Cursor says the combination gives it access to SpaceX’s GPU infrastructure to train stronger models at lower cost. The deal brings one of the leading AI coding platforms together with SpaceX’s rapidly expanding compute operation.

• Cursor is now a part of SpaceX · Cursor

• Form 8-K, 14 August 2026 · SpaceX

Apple trained its own AI model for China with Alibaba’s support.

Reuters reported that Apple has developed a China-specific LLM with Alibaba, alongside plans to integrate Qwen and other local technology into Apple Intelligence. China’s regulator registered Apple Intelligence in July, clearing the way for launch. If deployed as planned, Apple would become the first foreign company to offer its own proprietary AI model in China, rather than relying entirely on domestic models.

• Apple trains its own AI model for the China market with Alibaba’s support · Reuters

OpenAI’s enterprise business has overtaken consumer revenue.

CFO Sarah Friar told investors that enterprise customers rose 32% in July and now generate more revenue than OpenAI’s consumer business. The crossover shows how quickly the company is shifting from a consumer chatbot-led business towards selling AI as core enterprise infrastructure.

• OpenAI CFO Friar tells investors that enterprise is bigger than consumer · CNBC

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs trace each story to its original source and show how the week unfolded.

The week in five lines:

• Demis Hassabis stepped down as Google DeepMind CEO.

• UK AISI tests recorded 19 unsanctioned model actions against real targets.

• Google lined up $200bn to fund Anthropic’s chips.

• The EU AI Act obligations became enforceable.

• Stanford and Arc scientists built 16 AI-designed bacteriophages.

Browse by day:

• Weekend: OpenAI Astra · EU AI Act

• Monday: Qwen3.8-Max · US AI framework · Olix

• Tuesday: UK AISI · Google-Anthropic $200bn · Anthropic-Volta

• Wednesday: DeepMind shake-up · Jeff Dean · Meta Muse Code

• Thursday: AI-designed bacteriophages

• Friday: ByteDance 10T · Kimi K3 escape · Anthropic biology safeguards

Weekend (1-2 August)

OpenAI said its Astra model produced results on ten open problems in mathematics, quantum complexity and theoretical computer science.

OpenAI published a 249-page paper sharing a selection of ten results, each of which resolves or makes progress on long-standing open problems, together with machine-checkable proofs on GitHub. The results came at compute costs of c.$2,000 at Sol API rates. OpenAI showed the same model family privately to US policymakers days earlier.

• Ten advances in mathematics and theoretical computer science · OpenAI

EU AI Act obligations for general-purpose models became enforceable.

From 2 August the European Commission’s AI Office can investigate providers of general-purpose AI models, demand access for evaluations, order risk mitigations and fine up to 3% of global turnover. Separate transparency rules took effect requiring chatbots to disclose they are AI and synthetic images, audio and video to be labelled.

• Commission starts enforcing AI Act & transparency rules · European Commission

Monday 3 August

Alibaba said its 2.4-trillion-parameter Qwen3.8-Max tops Moonshot’s Kimi K3 and will publish the weights of Qwen3.8-Max and Qwen3.8-27B next week.

The model activates only 95B parameters per query through a mixture-of-experts design and carries a 1M-token context window. Open weights at that scale would be among the largest sets of frontier weights released.

• Qwen3.8-Max: a new bar for coding and cowork · Alibaba Cloud

• Alibaba adds to China AI breakthroughs with new Qwen model · Bloomberg

The White House said it met its deadline to establish a voluntary framework for evaluating advanced AI models, without publishing what it contains.

Details are being provided only to the companies taking part. Axios reported the next day that the text defined a covered frontier model as closed-source with state-of-the-art capabilities and national-security risks, leaving open-weights outside the perimeter.

• White House AI framework & Inside Trump’s AI framework · Axios · Axios

London chip startup Olix raised $312m led by Fundomo at a $3.3bn valuation, with Arm and the UK’s sovereign AI fund among the backers.

Up from just over $1bn when it raised $220m in February, a tripling inside six months. Olix is building optical AI-inference silicon aimed at Nvidia’s territory. Arm’s investment secures it a strategic foothold in a rising competitor’s hardware roadmap.

• OLIX raises $312m at a $3.3bn valuation & appoints Nick McKeown to board · OLIX

• UK AI chip startup Olix raises $312m at a $3.3bn valuation · Financial Times

Tuesday 4 August

The UK AI Security Institute recorded 19 unsanctioned actions against real people and organisations.

During a cyber evaluation between 25 and 28 July, the Institute found that agents under test took 19 unsanctioned actions against real people and organisations: 17 by Anthropic’s Mythos 5 and two by OpenAI’s GPT-5.6 Sol with its safety filters disabled.

One agent tried to insert malicious code into a public open-source project and used fake identities to socially engineer a maintainer to approve it. The reviewer refused the code, the AISI isolated its systems, notified GitHub, and no real-world harm resulted.

• Incident report: unsanctioned agent behaviour during testing · AI Security Institute

Google assembled a $200bn financing programme for Anthropic.

Google has arranged $200bn of financing to supply Anthropic, >$150bn of it tied to its own TPUs, together with Broadcom, Blackstone, Apollo and Morgan Stanley. A vehicle will buy the chips and lease them to Anthropic, moving it off Google’s balance sheet. This deepens Anthropic’s dependence on Google rather than Nvidia silicon.

• Inside Google’s $200bn Wall Street machine for Anthropic · Financial Times

Anthropic signed a $10bn, six-year deal for computing capacity with Volta Infra, an Nvidia-backed cloud startup founded only this year.

The capacity sits at a 133 MW site in Norway operated by crypto miner Bitdeer, running on hydroelectric power and Nvidia’s Vera Rubin chips, with delivery phased into early 2027. It puts a sizeable share of Anthropic’s compute outside the US.

• Volta launches: the utility of compute · Volta

• Anthropic signs a $10bn deal with cloud startup Volta · TechCrunch

SpaceX reported Q2 capital spending of $18.4bn, with $15.8bn of it going to its AI segment, and booked a $1.3bn AI operating loss.

The figure puts SpaceX among the largest single-quarter AI-infrastructure spenders anywhere, and SPCX fell >7% after hours. The next day Elon Musk said SpaceX will build exclusively on Nvidia, citing the Vera Rubin architecture, one of the larger single-buyer commitments Nvidia has won this year.

• SpaceX reports second quarter 2026 results · SpaceX

• SpaceX and Nvidia are taking their relationship exclusive · Business Insider

An internal TikTok document shows it withheld a safety-tuned version of its recommendation algorithm from 10% of US users as a control group.

One of those users was a 16-year-old who was served self-harm content and later died by suicide. A deliberate, documented withholding of a safety intervention is evidence that may drive litigation and legislation.

• TikTok withheld a safety feature from millions. One died by suicide · Bloomberg

Wednesday 5 August

Demis Hassabis stepped down as CEO of Google DeepMind.

Hassabis becomes chair of Deepmind and Alphabet chief scientist, allowing him to “put his full attention on actively shaping the future of AGI” and will lean into to his role at Isomorphic, noting “It’s time for AI to prove its unequivocal value to the world, and what better way to demonstrate that than to help finally cure diseases like cancer”.

CTO Koray Kavukcuoglu takes over day-to-day leadership as SVP, reporting to Sundar Pichai. Hassabis has run DeepMind since co-founding it in London. GOOG fell >3% on the news. Google has been moving AI decision-making towards Mountain View.

• The next chapter of our AI momentum · Google

Jeff Dean is leaving Google as chief scientist to co-found Discovery Loop, a startup chasing AI-driven breakthroughs in drug discovery and chip design.

Leaving after 27 years, he is joined by longtime collaborator Sanjay Ghemawat. Google is a founding investor and cloud partner in Discovery Loop. The move brings two of AI’s most influential pioneers further into life sciences.

• The next chapter of our AI momentum · Google

• Four top Google AI researchers form new startup · The New York Times

Meta released Muse Code in beta, a terminal coding agent running on a new model, Muse Spark 1.2.

Muse Spark 1.2 is priced at $1.25/m input tokens and $4.25/m output, undercutting coding offerings from Anthropic and OpenAI. The agent installs from the terminal and works across large repos with background sub-agents and a crash-safe event log.

• Introducing Muse Code and Muse Spark 1.2 · Meta AI Research

A filing shows Microsoft recorded $24.1bn from commercial arrangements with OpenAI, including revenue-sharing payments, in the year ended June.

Bloomberg estimated it represented more than half of Microsoft’s AI sales. The concentration cuts both ways: it quantifies how much of Microsoft’s AI business rests on one customer, and how much of OpenAI’s spending flows back to its own investor.

• Form 10-K for the fiscal year ended 30 June 2026 · Microsoft

• Microsoft’s AI sales mostly come from OpenAI, disclosures show · Bloomberg

Thursday 6 August

Stanford and Arc Institute scientists designed 16 viable bacteriophages with AI.

Writing in Science, the team used the Evo genome language model to generate whole bacteriophage genomes letter by letter, synthesised hundreds of candidates and found 16 that were viable. All 16 infect E. coli and pose no threat to humans. A cocktail of them overcame bacteria that had resisted a natural phage. The research highlights the need for biosecurity governance guardrails to be properly established.

• Generative design of bacteriophages with genome language models · Science

Two more frontier models breached real systems during safety testing.

OpenAI said the agents behind an earlier Hugging Face breach had built a covert message board to coordinate, sharing exploits through a package manager for months without anyone noticing. Separately, Meta’s Muse Spark 1.1 broke into another company after its evaluation partner Irregular left a sandbox misconfigured. In both cases the escape traced to the test harness, not the model, but a frontier system still ran an unsanctioned intrusion against a live target.

• OpenAI’s agents used a message board to plan their hacking spree · Wired

• A Meta AI model hacked another company during testing · The Information

Friday 7 August

ByteDance is pretraining a model with up to 10 trillion parameters.

Roughly three times the size of Moonshot’s Kimi K3 and above the 8T estimate for Anthropic’s Mythos 5, would make it one of the largest reported pretraining runs. Whether it yields a usable competitive frontier model is a separate question.

• ByteDance pretrains a model of up to 10T parameters · Financial Times

Moonshot’s Kimi K3 got outside its sandbox during a defensive cybersecurity evaluation and read the answers to the test.

Frontier Security ran the evaluation on the UK AI Security Institute (AISI) benchmark. It found the sandbox had left outbound HTTPS and DNS open. The model probed the settings, resolved github.com, cloned the benchmark’s own repository and read the solutions off disk rather than solving the tasks. Not a hack but specification gaming.

• Chinese model Kimi K3 breaks UK AISI benchmark evaluations · Frontier Security

• One of China’s most powerful AI models has also escaped containment · Wired

Anthropic loosened Claude Fable 5’s biology safeguards, cutting biology-related fallbacks by about 85%.

The company retrained the classifier that sits on biosecurity-relevant queries to cut false positives on everyday health, educational and clinical questions, while keeping blocks on dual-use areas such as virology and toxicology. Anyone using the model for life-sciences work should expect noticeably different behaviour.

• Improving Fable 5’s biology safeguards · Anthropic

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs trace each story to its original source and show how the week unfolded.

The week in five lines:

• Moonshot released Kimi K3 as open weights.

• Anthropic found three of its models had breached outside organisations.

• Google DeepMind introduced Gemini Robotics 2.

• Nvidia is in talks to backstop $250bn for OpenAI.

• Hugging Face mapped the agent that hacked it.

Browse by day:

• Weekend: chatbot bioweapon answers

• Monday: Kimi K3 · Nvidia-OpenAI · CXMT · Hugging Face breach

• Tuesday: Claude cryptanalysis · Pacing the Frontier · Amazon Nova

• Wednesday: Big Tech earnings · Moonshot raise

• Thursday: Gemini Robotics 2 · Anthropic breaches · GPT-5.6

• Friday: DeepSeek V4 Flash

Weekend (25-26 July)

Users have been coaxing chatbots into giving accurate guidance on mass-casualty attacks and bioweapons, the Wall Street Journal reported.

Staff at AI labs are among its sources, which allege that safeguards have failed in practice and answers to highly sensitive material come back correct rather than as garbled refusals. Biosecurity policy has been written around this failure mode and the burden is back on the labs to show their mitigations hold.

• Chatbots coaxed into bioweapon guidance · The Wall Street Journal

Monday 27 July

Moonshot released Kimi K3 as open weights.

Kimi K3 is Moonshot’s most capable model to date: a native multimodal, agentic mixture-of-experts system of 2.8T parameters (104B active), released under a bespoke license. Moonshot reports scores of 93.5 on GPQA Diamond and 88.3 on Terminal-Bench, at maximum reasoning. Releasing weights on this scale publicly puts the China lab at the centre of an intense US debate over whether to ban foreign AI models.

• Kimi K3 model card · Hugging Face

Nvidia is in talks to backstop $250bn of OpenAI financing.

The Wall Street Journal reported that Nvidia was in talks to guarantee roughly $250bn of financing for OpenAI, tied to a 10 GW data centre SoftBank is developing in Ohio. Separately, Bloomberg put Nvidia’s cheque to Ilya Sutskever’s Safe Superintelligence at about $5bn, larger than everything the lab had raised before.

• Nvidia to guarantee OpenAI financing · The Wall Street Journal

• Nvidia to invest $5bn in Safe Superintelligence · Bloomberg

Anthropic set out its position on open-weights models, with Dario Amodei writing that the company has never advocated a ban.

Amodei calls for chip-export controls, a crackdown on industrial-scale distillation, and mandatory safety testing for sufficiently capable models, open or closed. A direct answer to claims Anthropic had been lobbying against open models.

• Anthropic’s position on open weights · Anthropic

CXMT closed 466% above its offer price on its Shanghai debut, making it the most valuable China-listed company.

The listing raised 57.9bn yuan ($8.6bn). Memory is a key constraint for AI servers right now and the listing gives China’s largest DRAM maker capital to serve this demand.

• CXMT surges on Shanghai debut · CNBC

Hugging Face traced the OpenAI agent that breached it.

Hugging Face published a step-by-step forensic timeline of the July intrusion, reconstructing about 17,600 actions the rogue OpenAI agent took across a 4.5 day window from 9-13 July: an initial sandbox escape, lateral movement through Kubernetes clusters on stolen service-account tokens, and data exfiltration. The analysis used open-weight GLM-5.2 after Claude models declined on safety grounds.

• Technical timeline of the agent intrusion · Hugging Face

Tuesday 28 July

More than 1,100 staff at AI companies signed a letter asking the US government to help pace AI development, Bloomberg reported.

Signatories include OpenAI’s Jakub Pachocki and Thinking Machines’ John Schulman, putting research leaders on record asking for constraints on their own labs. OpenAI and Anthropic later attached themselves to the effort, branded Pacing the Frontier .

• Statement from 1,367 employees of AI companies · Pacing the Frontier

• Staff letter asks Washington to pace AI · Bloomberg

Wednesday 29 July

Google DeepMind has broken up the team behind AlphaFold, the FT reported, reassigning most of the original authors.

Roughly a quarter of the authors have now left DeepMin, including John Jumper, who shared the 2024 Nobel Prize with Demis Hassabis and David Baker and has left to join Anthropic. This raises questions about the direction of DeepMind’s scientific mission.

• GDM dismantles Nobel-winning AlphaFold team in strategy shift · Financial Times

Meta and Microsoft quarterly results revealed rising AI capex.

Meta’s Q2 revenue rose 28% YoY to $60.8bn, but free cash flow fell to $784m as capital spending climbed, and it lifted its full-year expense outlook to $165bn-$169bn. Microsoft’s Q4 revenue rose 18% to $90bn with net income up 31% to $35.8bn, and it said Azure revenue passed $100bn for the first time, with Azure up 43% in the quarter.

• Meta second-quarter results · Meta

• Microsoft fiscal fourth-quarter results · Microsoft

Moonshot AI closed a $3.5bn round at a $35bn valuation, Bloomberg reported, well past its $1bn-$2bn target, and plans a Hong Kong IPO as soon as 2026.

Raising double its target ahead of an IPO demonstrates that non-US frontier labs can mobilise significant amounts of capital independent of American venture networks.

• Moonshot passes funding goal · Bloomberg

Thursday 30 July

Google DeepMind introduced Gemini Robotics 2.

DeepMind describes it as three models working together: a vision-language-action (VLA) controller, an embodied-reasoning planner, and an on-device variant that adapts to a new robot body in a few hours. Together they give intelligent whole-body control across humanoid and bi-arm robots, including the Apptronik Apollo 2, and support multi-robot collaboration. Available in AI Studio and Gemini Enterprise Agent Platform, the VLA and On-Device models are available to early-access partners.

• Gemini Robotics 2 brings whole body intelligence to robots · Google DeepMind

Anthropic found three of its models had breached outside organisations.

Anthropic said a review after the OpenAI-Hugging Face incident turned up three cases where Claude reached real systems it should not have: extracting credentials and production data from one company, publishing malicious code to PyPI that hit 15 real systems, and compromising an internet-facing application. The models involved were Opus 4.7, Mythos 5 and an internal research model. The company framed it as a harness and operational failure. Containment failures were not unique to one lab.

• Investigating three real-world incidents in our cybersecurity evaluations · Anthropic

OpenAI announced price cuts for two of its GPT-5.6 models.

The company said it was lowering the price of GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20% after efficiency gains in serving. OpenAI is now competing directly on cost per unit of intelligence, with major implications for enterprise inference budgets.

• Advancing the price-performance frontier with GPT‑5.6 · OpenAI

AI hedge fund Situational Awareness sold the bulk of its stock portfolio to Citadel after deep losses, the Wall Street Journal reported.

Leopold Aschenbrenner’s fund was among the most concentrated bets on the AI trade. Its assets fell to about $10bn after it liquidated positions to meet margin calls, and its forced unwind is now itself a factor in AI equity prices.

• Citadel buys Situational Awareness portfolio after AI losses · The Wall Street Journal

Friday 31 July

DeepSeek released its V4 Flash model in public beta.

The company said the update improves agent capabilities over the V4 Pro Preview while keeping the same architecture, re-post-trained rather than rebuilt. Artificial Analysis scored it 52 on its Intelligence Index and classifies it as a large open-weights model, at $0.14 per million input tokens and $0.28 per million output tokens.

• DeepSeek V4 Flash release notes · DeepSeek

• V4 Flash intelligence and price analysis · Artificial Analysis

Chinese military researchers distilled OpenAI and Anthropic models to train domestic systems for defence work, Reuters reported.

Distillation uses a frontier model’s outputs as synthetic training data to instruct smaller student models, side-stepping the massive pre-training compute targeted by US export controls. The findings coincided with state media reporting a fresh call from Xi Jinping for the military to accelerate its deployment of AI and autonomous systems.

• Chinese military researchers tap US AI models to train defence systems · Reuters

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

• Cerebras filed for a Nasdaq listing after 2025 revenue rose 76% to $510m.

• Anthropic shipped Claude Opus 4.7 with its cyber skills deliberately reduced.

• The UK’s AI Security Institute saw Mythos complete a 32-step network attack.

• OpenAI answered Mythos with GPT-5.4-Cyber, released to vetted defenders only.

• GPT-5.4 Pro solved Erdős problem #1196, and Terence Tao praised its method.

Browse by day:

• Weekend: Berkeley benchmarks

• Monday: OpenAI London · AISI · Erdős

• Tuesday: Nscale · GPT-5.4-Cyber · Meta-Broadcom · Gemini Robotics

• Wednesday: Jane Street · WhatsApp

• Thursday: Qwen · Opus 4.7 · Codex · Callosum · Rosalind · π0.7 · London

• Friday: Claude Design · Cerebras · Kevin Weil

Weekend (11-12 April)

Researchers at UC Berkeley showed that eight leading AI agent benchmarks can be gamed to near-perfect scores without solving a single task.

Their agent found that a 10-line Python file marks every SWE-bench Verified instance resolved, and that a fake curl wrapper scores 100% on all 89 Terminal-Bench tasks.

• How We Broke Top AI Agent Benchmarks · Berkeley RDI

Monday 13 April

OpenAI signed a lease for its first permanent London office, with room for >500 staff.

The 88,500 sq ft office is in King’s Cross, and OpenAI employs about 200 people in London. The lease came days after OpenAI paused its Stargate UK project.

• OpenAI announces first permanent London office · CNBC

The UK AI Security Institute saw Claude Mythos Preview complete a 32-step simulated network attack.

Anthropic’s withheld model is the first to complete The Last Ones, AISI’s 32-step simulated attack that runs from reconnaissance to full network takeover and takes a human ~20 hours. It finished in 3 of 10 attempts and averaged 22 steps, while Claude Opus 4.6, the next best, averaged 16. On expert-level capture-the-flag tasks, which no model could complete before April 2025, it succeeds 73% of the time. AISI directed the model and gave it network access, with a budget of up to 100M tokens per run.

• Our evaluation of Claude Mythos Preview’s cyber capabilities · AI Security Institute

GPT-5.4 Pro (OpenAI) solved Erdős problem #1196, a conjecture about primitive sets.

Liam Price posted the proof in the Erdős Problems forum on Monday. Terence Tao wrote on Tuesday that its method would be “a meaningful contribution to the anatomy of integers that goes well beyond the solution of this particular Erdos problem”. Math, Inc. had formally verified the proof by Thursday.

• 1196 Discussion Thread · Erdős Problems

Tuesday 14 April

Microsoft will take >30,000 Nvidia Rubin GPUs from Nscale at a Norwegian site once earmarked for OpenAI’s Stargate.

The GPUs go into Nscale’s 230 MW campus at Narvik, with capacity arriving in 2027. OpenAI failed to conclude an agreement for the site, Bloomberg reported.

• Nscale announces expanded deal with Microsoft in Norway · Nscale

• Microsoft contracts 30,000 GPUs from Nscale at Norway data center · DCD

OpenAI released GPT-5.4-Cyber to vetted security defenders.

GPT-5.4-Cyber is a version of GPT-5.4 fine-tuned for cyber work, with fewer restrictions and a lower refusal boundary for legitimate security tasks. It adds binary reverse engineering, so analysts can examine compiled software for malware and vulnerabilities without the source code. OpenAI is starting with a limited deployment to vetted security users in the top tiers of its Trusted Access for Cyber programme.

• Trusted access for the next era of cyber defense · OpenAI

Meta committed to >1 GW of custom MTIA chips co-developed with Broadcom, and Broadcom chief executive Hock Tan will leave Meta’s board.

The 1 GW is the first phase of a multi-gigawatt roll-out across several chip generations. Broadcom says the MTIA will be the industry’s first 2nm AI accelerator. Tan becomes an adviser to Meta.

• Meta Partners With Broadcom to Co-Develop Custom AI Silicon · Meta

• Broadcom Announces Extended Partnership with Meta · Broadcom

Google DeepMind released Gemini Robotics-ER 1.6, a reasoning model that plans robot tasks, through the Gemini API.

DeepMind says it beats Gemini Robotics-ER 1.5 and Gemini 3.0 Flash on spatial reasoning such as pointing, counting and success detection. It can also read gauges and sight glasses, a capability developed with Boston Dynamics.

• Gemini Robotics-ER 1.6: Enhanced Embodied Reasoning · Google DeepMind

Wednesday 15 April

Jane Street committed about $6bn to CoreWeave’s AI cloud, including Nvidia Vera Rubin capacity, and bought $1bn of CoreWeave stock.

It paid $109 a share for the Class A stock.

• Jane Street Signs $6 Billion AI Cloud Agreement With CoreWeave · CoreWeave

The European Commission told Meta it intends to order WhatsApp reopened to rival AI assistants on the terms in place before October 2025.

The Commission’s supplementary statement of objections says Meta’s March switch to paid access is in effect equivalent to its earlier ban. It also extended the investigation to Italy, so it now covers the whole EEA.

• Commission sends Meta fresh WhatsApp charge sheet · European Commission

Thursday 16 April

Alibaba’s Qwen team released Qwen3.6-35B-A3B, the first open-weight Qwen3.6 model.

It has 35B parameters, 3B active. Alibaba says it scores 73.4 on SWE-bench Verified and 51.5 on Terminal-Bench 2.0, against 52.0 and 42.9 for Gemma 4 31B (Google).

• Qwen/Qwen3.6-35B-A3B · Hugging Face

Anthropic released Claude Opus 4.7 and said it had deliberately held back its cyber capabilities.

Anthropic calls the model a notable improvement on Claude Opus 4.6 in advanced software engineering, with the biggest gains on the hardest tasks. It is less broadly capable than Claude Mythos Preview, and has reduced cyber capabilities, shipping with safeguards that block high-risk cybersecurity requests.

Prices stay at $5/m input and $25/m output tokens. Artificial Analysis scores it 57.3 on its Intelligence Index, tied for first with Gemini 3.1 Pro at 57.2 and GPT-5.4 at 56.8.

• Introducing Claude Opus 4.7 · Anthropic

• Opus 4.7: Everything you need to know · Artificial Analysis

OpenAI gave Codex control of the user’s computer, an in-app browser, image generation, memory and >90 new plugins.

Codex, used weekly by >3 million developers, can now schedule its own work and resume long tasks. Computer use starts on macOS and reaches EU and UK users later.

• Codex for (almost) everything · OpenAI

The UK’s £500m Sovereign AI Unit made its first equity investment, in Callosum, which builds software that makes different kinds of chip work together.

Six more start-ups get up to 1 million GPU hours each on the UK’s supercomputers. Technology Secretary Liz Kendall announced the awards.

• AI firms get first backing through UK’s Sovereign AI · GOV.UK

The White House budget office began work on safeguards before possibly giving federal agencies a modified Claude Mythos Preview.

Federal chief information officer Gregory Barbaccia’s Tuesday email said OMB is working on safeguards before “potentially releasing a modified version of the model to agencies”, Bloomberg and Nextgov/FCW reported. On Friday an OMB spokesperson said “OMB is not giving access to anything to agencies”.

• OMB is ‘not giving access to anything to agencies,’ official says · Nextgov/FCW

OpenAI launched GPT-Rosalind, a reasoning model for research, drug discovery and translational medicine, as a research preview for vetted customers.

Amgen, Moderna, the Allen Institute and Thermo Fisher Scientific are among the first users, through OpenAI’s trusted access programme, Axios reported.

• Introducing GPT-Rosalind for life sciences research · OpenAI

• OpenAI launches new AI model for life sciences research · Axios

Physical Intelligence released π0.7, a robot model it says can recombine skills learned in training to do tasks it was never taught.

The company says the model matched its own task-specific models at making coffee, folding laundry and assembling boxes. Coached step by step in plain language, it handled an air fryer far better. The paper claims only “early signs” of generalisation.

• A Steerable Model with Emergent Capabilities · Physical Intelligence

• Physical Intelligence says its robot brain can do tasks never taught · TechCrunch

Anthropic said it was taking office space for 800 people in London’s Knowledge Quarter, days after OpenAI’s London lease.

It has >200 staff in the city.

• Anthropic unveils plans for major London expansion · CNBC

Friday 17 April

Anthropic launched Claude Design, a Labs product for making prototypes, slides and one-pagers with Claude - Figma and Adobe shares fell.

Claude Design runs on Claude Opus 4.7 and is in research preview for Pro, Max, Team and Enterprise subscribers. Work exports to Canva, PDF, PPTX or HTML.

• Introducing Claude Design by Anthropic Labs · Anthropic

• Anthropic launches “Claude Design,” sending Figma shares down · Sherwood News

Cerebras filed to list on Nasdaq after revenue rose 76% in 2025.

Cerebras’s S-1 shows 2025 revenue of $510m and net income of $238m, against a $482m net loss in 2024. It will list under the ticker CBRS, with Morgan Stanley, Citigroup, Barclays and UBS among the lead underwriters. Two UAE customers dominate. Mohamed bin Zayed University of Artificial Intelligence was 62% of 2025 revenue and G42 24%. OpenAI has lent Cerebras $1bn and holds a warrant to buy its stock, and in January agreed a multi-year deal worth >$20bn to deploy 750 MW of Cerebras compute. The Information reported an indicative valuation of >$35bn.

• Cerebras - S-1 (April 2026) · SEC

• Cerebras Prepares Public Listing, Eyes $35 Billion-Plus Valuation · The Information

Kevin Weil left OpenAI, which is splitting up his OpenAI for Science group and closing Prism, its AI workspace for scientists.

Prism’s team of about 10 moves under Codex chief Thibault Sottiaux. Bill Peebles, who led Sora, also announced his departure on Friday.

• Today is my last day at OpenAI · Kevin Weil on X

• OpenAI Executive Kevin Weil Is Leaving the Company · WIRED

• Kevin Weil and Bill Peebles exit OpenAI as company sheds ‘side quests’ · TechCrunch

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

• Anthropic withheld Mythos, which found thousands of flaws in major software.

• Anthropic signed for gigawatts of Google TPUs as run-rate revenue passed $30bn.

• Meta launched Muse Spark, its first model to ship without open weights.

• OpenAI paused Stargate UK over energy costs, and Stargate leaders left for Meta.

• The US Treasury and Fed called in bank chiefs over the cyber risks of Mythos.

Browse by day:

• Monday: Google-Broadcom TPUs · Samsung · OpenAI policy

• Tuesday: Mythos · Musk suit · GLM-5.1 · Terafab

• Wednesday: Muse Spark · Stargate UK · Anthropic appeal · xAI · Managed Agents

• Thursday: Banks · CoreWeave · Jassy · Colorado · Emil Michael · Illinois

• Friday: TSMC · HappyHorse · OpenAI Mac apps

Monday 6 April

Anthropic signed for multiple gigawatts of Google TPU capacity and said its run-rate revenue has passed $30bn.

The capacity starts coming online in 2027, mostly in the US. Run-rate revenue was about $9bn at the end of 2025. More than 1,000 business customers now spend >$1m a year each, double the number in February. Broadcom will develop and supply Google’s TPUs under a long-term agreement, and supply networking and rack components through to 2031. Anthropic will take 3.5 GW of the capacity through Broadcom. How much it uses depends on Anthropic’s continued commercial success.

• Anthropic expands partnership with Google and Broadcom · Anthropic

• Broadcom Inc. Form 8-K, 6 April 2026 · SEC

Samsung Electronics guided to Q1 2026 operating profit of about 57.2 trillion Korean won, roughly 8.5 times the 6.7 trillion won of Q1 2025.

Sales came in at about 133 trillion won, up about 68% YoY from 79.1 trillion won. The surge was driven by booming AI data-centre demand, which has pushed memory chip prices sharply higher as manufacturers divert capacity towards advanced AI chips, squeezing supplies of conventional DRAM and NAND.

• Samsung Announces Earnings Guidance for First Quarter 2026 · Samsung

OpenAI published a policy paper on preparing for superintelligence, proposing a public wealth fund, higher capital gains taxes at the top and a four-day week.

Employers and unions would pilot a 32-hour or four-day week at full pay, made permanent if productivity holds up.

• Industrial policy for the Intelligence Age: Ideas to Keep People First · OpenAI

Tuesday 7 April

Anthropic withheld Claude Mythos Preview from general release because of its hacking ability.

Mythos has found thousands of high-severity vulnerabilities, including in every major operating system and web browser. They include a 27-year-old OpenBSD flaw that let an attacker crash any machine just by connecting to it, and a 16-year-old bug in FFmpeg. Anthropic does not plan to make Mythos generally available and formed Project Glasswing instead. AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan, Linux Foundation, Microsoft, Nvidia and Palo Alto Networks will use it for defensive work, and >40 other organisations that maintain critical software get access.

On Anthropic’s figures, Claude Mythos Preview scores 77.8% on SWE-Bench Pro against 53.4% for Claude Opus 4.6, and 93.9% against 80.8% on SWE-Bench Verified. The system card records that the model, told to escape a sandbox, did so and emailed a researcher. It then posted details of its exploit to public-facing websites unasked.

• Project Glasswing: Securing critical software for the AI era · Anthropic

• System Card: Claude Mythos Preview · Anthropic

Elon Musk told the court he will seek no damages for himself from OpenAI, and wants any ill-gotten gains returned to its charity.

He will also ask for Sam Altman’s removal from the nonprofit’s board, and for Altman and Greg Brockman to be removed as officers of the for-profit company.

• Amended notice of remedies, 7 April · CourtListener

Z.ai released GLM-5.1, a 754B-parameter mixture of experts (MoE) model, as open source.

On Z.ai’s figures, it scores 58.4 on SWE-Bench Pro, ahead of GPT-5.4 (OpenAI) at 57.7, Claude Opus 4.6 at 57.3 and Gemini 3.1 Pro (Google) at 54.2. Z.ai built it for long agentic work. In one demonstration it kept improving its own output for 8 hours.

• GLM-5.1: Towards Long-Horizon Tasks · Z.ai

• GLM ships 5.1 open source LLM, beating Opus 4.6 and GPT-5.4 · VentureBeat

Intel joined Elon Musk’s Terafab chip project alongside SpaceX, Tesla and xAI.

Intel said its ability to design, fabricate and package chips at scale will help Terafab towards its goal of 1 TW of compute a year.

• Intel is proud to join the Terafab project · Intel on X

Wednesday 8 April

Meta released Muse Spark, the first model from Meta Superintelligence Labs.

Muse Spark is a multimodal reasoning model and the first in the Muse family. It runs in meta.ai and the Meta AI app, and selected users get a private API preview. Artificial Analysis scores it 52 on its Intelligence Index, behind only Gemini 3.1 Pro (Google), GPT-5.4 (OpenAI) and Claude Opus 4.6 (Anthropic). It is Meta’s first release without open weights, though Meta plans an open-source version.

• Introducing Muse Spark: Scaling Towards Personal Superintelligence · AI at Meta

• Muse Spark: Meta is back in the AI race · Artificial Analysis

• Meta debuts Muse Spark, first AI model under Alexandr Wang · Axios

OpenAI paused Stargate UK, citing energy costs and regulation.

OpenAI announced the project in September with Nscale and Nvidia, planning up to 8,000 GPUs and potentially 31,000 over time. Politico reported the pause on Wednesday. OpenAI told CNBC it will move forward “when the right conditions such as regulation and the cost of energy enable long-term infrastructure investment”. Three leaders of OpenAI’s data centre effort, Peter Hoeschele, Shamez Hemani and Anuj Saharan, have left and are joining Meta, Bloomberg reported on Friday.

• OpenAI halts UK stargate project amid regulatory, energy price concerns · CNBC

• Former OpenAI Stargate Leaders Plan to Join Meta Platforms · Bloomberg Law

A US appeals court refused to pause the Pentagon’s supply chain risk label on Anthropic while its challenge proceeds.

Judges Henderson, Katsas and Rao of the DC Circuit denied the stay in Anthropic’s case against the US Department of War and set oral argument for 19 May.

• Order denying stay, 8 April · DC Circuit

xAI is overhauling its engineering team, and SpaceX’s Michael Nicolls told staff the company is “clearly behind”.

Nicolls, who runs Starlink at SpaceX, is now xAI president, Business Insider reported. His memo called the training performance of xAI’s compute “embarrassingly low”.

• XAI Reorganizes Its Engineering Team Ahead of SpaceX IPO · Business Insider

Anthropic launched Claude Managed Agents, a hosted service for building and running AI agents, in public beta.

It costs standard Claude Platform token rates plus $0.08 per session-hour of active runtime. Notion and Rakuten are early users.

• Claude Managed Agents: get to production 10x faster · Anthropic

Thursday 9 April

The US Treasury and Federal Reserve called bank chiefs to a meeting on the cyber risks of Mythos.

They included Citigroup’s Jane Fraser, Morgan Stanley’s Ted Pick, Bank of America’s Brian Moynihan, Wells Fargo’s Charlie Scharf and Goldman Sachs’ David Solomon. Bloomberg reported that Goldman Sachs, Citigroup, Bank of America and Morgan Stanley are testing Mythos internally. JPMorgan Chase is the only bank named in Project Glasswing. The week before Mythos launched, Vice President JD Vance and Bessent had questioned tech chief executives including Dario Amodei, Elon Musk, Sundar Pichai, Sam Altman and Satya Nadella on AI model security.

• Bessent Urgently Summons Bank CEOs Over Anthropic’s New AI · Bloomberg Law

• Wall Street Banks Try Out Anthropic’s Mythos as US Urges Testing · Bloomberg Law

• Trump administration met with tech giants before Mythos release · CNBC

CoreWeave signed a $21bn AI cloud deal with Meta, and a day later a multi-year compute deal with Anthropic.

Meta’s capacity runs to December 2032 and includes some of the first deployments of Nvidia’s Vera Rubin platform. Anthropic’s starts coming online later this year. CoreWeave says nine of the ten leading model providers now use its platform.

• CoreWeave and Meta Announce $21 Billion Infrastructure Agreement · CoreWeave

• CoreWeave Announces Multi-Year Agreement With Anthropic · CoreWeave

Andy Jassy’s shareholder letter put AWS’s AI revenue at a $15bn annual run rate.

Amazon’s chips business runs at >$20bn a year. Jassy defended about $200bn of capex: “We’re not investing approximately $200 billion in capex in 2026 on a hunch”.

• Amazon CEO Andy Jassy’s 2025 Letter to Shareholders · Amazon

xAI sued Colorado’s attorney general, Philip Weiser, in federal court over the state’s AI anti-discrimination law.

• X. AI LLC v. Weiser, 1:26-cv-01515 · CourtListener

• xAI sues Colorado over first state AI anti-discrimination law · Financial Times

Pentagon research chief Emil Michael sold his xAI stake for between $5m and $25m after the Pentagon struck deals with the company, the Guardian reported.

Michael had declared the stake at $500,000 to $1m in March 2025 and sold it on 9 January, ethics disclosures show. The Pentagon said he is “in full compliance with all ethics laws and regulations”.

• US official overseeing AI reaped millions after Pentagon agreement · The Guardian

OpenAI backed an Illinois bill that would shield frontier AI developers from liability for harms such as the death or serious injury of 100 or more people.

SB 3444 also covers at least $1bn of property damage, and applies to models trained at a computational cost of >$100m. Developers keep the shield if they did not act intentionally or recklessly and publish safety, security and transparency reports. OpenAI testified in favour of it.

• OpenAI Backs Bill That Would Limit Liability for AI-Enabled Mass Deaths · WIRED

Friday 10 April

TSMC’s Q1 2026 revenue rose 35.1% YoY to NT$1,134bn ($35.7bn).

March revenue alone was NT$415.2bn ($13.1bn), up 45.2% YoY.

• TSMC March 2026 Revenue Report · TSMC

Alibaba claimed HappyHorse-1.0, the anonymous video model that had gone to the top of Artificial Analysis’s leaderboards.

It leads the silent-video leaderboards and ties ByteDance’s Seedance 2.0 with audio. It comes from the Alibaba Token Hub group and is in closed beta, with an API due soon.

• Revealing HappyHorse-1.0 as Alibaba’s video model · Artificial Analysis on X

• Alibaba Unveils HappyHorse After It Tops Video Rankings · Caixin Global

OpenAI said a GitHub Actions workflow that signs its macOS apps ran a malicious version of the Axios library on 31 March.

It had access to signing material for ChatGPT Desktop, Codex, Codex-cli and Atlas. OpenAI found no evidence user data or software was compromised, but treats the certificate as compromised and is replacing it. Mac users must update by 8 May.

• Our response to the Axios developer tool compromise · OpenAI

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

• Anthropic accidentally published Claude Code’s internal source code on npm.

• Eli Lilly signed an AI drug deal with Insilico Medicine worth up to $2.75bn.

• Google released its Gemma 4 open models in four sizes under Apache 2.0.

• Baidu robotaxis stopped in the road across Wuhan after a system failure.

• CoreWeave closed an $8.5bn investment-grade loan secured by AI compute.

Browse by day:

• Weekend: Eli Lilly-Insilico

• Monday: Mistral · UK FRC · California · Copilot tips

• Tuesday: Claude Code · OpenAI · Marvell · CoreWeave · Perplexity · FreeBSD · Baidu

• Wednesday: Bank of England

• Thursday: Gemma 4 · Anthropic appeal · Coefficient · MAI · Qwen · Cursor · TBPN

• Friday: Microsoft Japan · Mercor · OpenClaw

Weekend (28-29 March)

Eli Lilly signed an AI drug discovery deal with Insilico Medicine worth up to $2.75bn.

Hong Kong-listed Insilico gets $115m upfront. Development, regulatory and commercial milestones make up the rest, plus tiered royalties. Lilly takes an exclusive worldwide licence to preclinical oral drug candidates. The two will also run R&D programmes on targets Lilly selects.

• Insilico Medicine Announces Global R&D Collaboration with Lilly · Insilico Medicine

Monday 30 March

Mistral AI raised $830m in its first debt financing, to run its own data centre near Paris.

The money finances 13,800 Nvidia GB300 GPUs at Bruyères-le-Châtel and brings its powered capacity to 44 MW. Seven banks lent it, including Bpifrance, BNP Paribas, HSBC and MUFG. Mistral is building 200 MW across Europe by the end of 2027.

• We’ve announced today our first debt financing · Mistral AI on LinkedIn

The UK’s Financial Reporting Council (FRC) issued AI guidance for audit firms, which it says is the first from any audit regulator.

The FRC says the human auditor stays accountable for audit quality, and the guidance does not change that.

• New guidance supports audit firm adoption of emerging AI technologies · FRC

California Governor Gavin Newsom signed an executive order raising the bar for AI companies that want state contracts.

The state’s Government Operations Agency will design contracting processes that vet companies on safeguards against illegal content, biased models and breaches of civil rights and free speech. The order also lets California separate its procurement authorisation from the federal government’s, and asks for watermarking guidance for AI-generated images and video.

• Newsom signs first-of-its-kind executive order on AI · Governor of California

GitHub Copilot inserted a promotional tip for itself and Raycast into a developer’s pull request, and GitHub removed agent tips the same day.

Zach Manson said a colleague had only asked Copilot to fix a typo. The Register found >11,400 pull requests carrying the same tip. GitHub’s Martin Woodward said GitHub does not plan to include advertisements, blamed a tip that “surfaced in the wrong context” and said agent tips have now been disabled from pull request comments.

• Copilot Edited an Ad Into My PR · Zach Manson

• GitHub backs down, kills Copilot PR ‘tips’ after backlash · The Register

Tuesday 31 March

Anthropic accidentally published internal Claude Code source code in an npm package.

Version 2.1.88 shipped with a source map that exposed the code. Security researcher Chaofan Shou pointed it out on X early on Tuesday. Anthropic called it “a release packaging issue caused by human error, not a security breach” and said no sensitive customer data or credentials were exposed. It came five days after a leaked draft revealed Claude Mythos.

Copies spread across GitHub. Anthropic’s copyright notice led GitHub to disable a network of about 8,100 repositories. A day later Anthropic withdrew the notice for all but one repository and 96 named forks, and asked GitHub to reinstate the rest.

• Claude code source code has been leaked via a map file · Chaofan Shou on X

• Anthropic leaks part of Claude Code’s internal source code · CNBC

• DMCA notice, 31 March · GitHub · Partial retraction, 1 April · GitHub

OpenAI closed its funding round at $122bn, $12bn more than it announced in February.

The round values OpenAI at $852bn after the new money. SoftBank co-led with a16z, D. E. Shaw Ventures, MGX, TPG and accounts advised by T. Rowe Price. For the first time OpenAI sold to individual investors through banks, raising >$3bn that way. OpenAI says it now generates $2bn of revenue a month. Enterprise makes up >40% of it and is on track to match consumer by the end of 2026.

• OpenAI raises $122 billion to accelerate the next phase of AI · OpenAI

• OpenAI closes funding round at an $852 billion valuation · CNBC

Over a hundred Baidu robotaxis stopped mid-traffic across Wuhan after what police called a system failure.

Wuhan traffic police said that from 20:57 local time on Tuesday they received calls about Apollo Go cars stopped in the road. They put it down to a system failure. Passengers got out safely, nobody was hurt and the cause is under investigation.

• Wuhan Apollo Go stoppage, with the police notice · The Paper

• Baidu’s Apollo Go self-driving cars stop mid-traffic in China · BBC News

Nvidia invested $2bn in Marvell, and the two will work together on silicon photonics.

Marvell joins Nvidia’s NVLink Fusion platform, supplying custom XPUs and compatible scale-up networking to customers building on Nvidia architectures.

• NVIDIA Ecosystem Expands as Marvell Joins Forces Through NVLink Fusion · Nvidia

CoreWeave closed an $8.5bn loan it calls the first investment-grade financing backed by AI compute.

It is secured by the compute and a customer contract. Moody’s rated it A3 and DBRS A (low). CoreWeave can draw about $7.5bn at first. Blackstone Credit & Insurance anchored the facility, which matures in March 2032.

• Landmark $8.5bn first investment-grade rated GPU-backed financing · CoreWeave

A class action filed in San Francisco accuses Perplexity of sharing users’ conversations with Meta and Google, even in Incognito mode.

The complaint says the trackers load as soon as users log in.

• Class action complaint, 31 March · CourtListener

Security firm Calif said Claude wrote a working remote exploit for a FreeBSD kernel flaw, delivering a root shell in about eight hours.

Calif calls it the first remote kernel exploit both discovered and exploited by an AI. FreeBSD fixed the flaw, CVE-2026-4747, on 26 March in an advisory crediting “Nicholas Carlini using Claude, Anthropic”.

• MAD Bugs: Claude Wrote a Full FreeBSD Remote Kernel RCE with Root Shell · Calif

• FreeBSD-SA-26:08.rpcsec_gss · FreeBSD

Wednesday 1 April

The Bank of England’s Financial Policy Committee said the Middle East conflict has added to the threats facing AI company valuations.

The committee said valuations of US AI-focused technology companies remain stretched, and a sharp correction could tighten funding for UK companies. Shortages of input chemicals and materials could also bottleneck the AI infrastructure build-out.

• Financial Policy Committee Record, April 2026 · Bank of England

Thursday 2 April

Google released Gemma 4, its open model family, under the Apache 2.0 licence.

The family comes in four sizes. E2B and E4B target phones and edge devices. A 26B mixture of experts model and a 31B model run offline on personal computers.

• Gemma 4: Byte for byte, the most capable open models · Google

The US government appealed the court order that blocked part of the Pentagon’s supply chain risk label on Anthropic.

Government lawyers asked the Ninth Circuit to overturn Judge Rita Lin’s preliminary injunction of 26 March. The appeals court set 30 April for the government’s opening brief and 28 May for Anthropic’s answer. Anthropic sued the Pentagon on 9 March.

• Notice of appeal · CourtListener · Ninth Circuit schedule notice · CourtListener

Anthropic acquired start-up Coefficient Bio for about $400m, The Information and Newcomer reported.

• Anthropic Acquires Startup Coefficient Bio for About $400 Million · The Information

• Anthropic Buys Stealth Coefficient Bio in $400M+ Stock Deal · Newcomer

Microsoft released three in-house models through its Foundry platform: MAI-Transcribe-1, MAI-Voice-1 and MAI-Image-2.

Microsoft is deploying them in its products. Transcription starts at $0.36 per hour, voice at $22/m characters, and images at $5/m input and $33/m image output tokens.

• Announcing 3 new world class MAI models, available in Foundry · Microsoft AI

Alibaba released Qwen3.6-Plus, a hosted model with a 1M-token context window aimed at agentic coding.

It is available through Alibaba Cloud Model Studio and works with coding agents including OpenClaw and Claude Code.

• Qwen3.6-Plus: Towards Real World Agents · Alibaba Cloud

Cursor launched Cursor 3, a new interface built from scratch for running many coding agents in parallel.

Local and cloud agents sit in one sidebar, and sessions move between the two. Users can switch back to the Cursor IDE at any time.

• Meet the new Cursor · Cursor

OpenAI acquired TBPN, the tech talk show, and promised it editorial independence.

TBPN will pick its own guests and programming. It sits in OpenAI’s strategy organisation, reporting to Chris Lehane. Two weeks earlier OpenAI agreed to buy Astral, maker of Python developer tools.

• OpenAI acquires TBPN: Accelerating the global conversation around AI. · OpenAI

Friday 3 April

Microsoft announced a $10bn investment in Japan to run from 2026 to 2029.

Prime Minister Sanae Takaichi called it Microsoft’s largest-ever investment in the country. Sakura Internet and SoftBank will sell GPU compute through Azure with data kept in Japan. Microsoft plans to train >1 million engineers and workers by 2030.

• Microsoft deepens its commitment to Japan with $10 billion investment · Microsoft

WIRED reported that Meta paused all its work with data vendor Mercor while it investigates a security breach there.

OpenAI has not stopped its projects but is investigating whether its training data was exposed. Mercor had said on Tuesday that it was one of thousands of companies hit by the compromise of the open-source LiteLLM project, as extortion group Lapsus$ claimed access to its data.

• Meta Pauses Work With Mercor After Data Breach Puts AI Secrets at Risk · WIRED

• Mercor says it was hit by cyberattack tied to compromise of LiteLLM · TechCrunch

Anthropic said Claude subscriptions would stop covering usage through third-party tools such as OpenClaw from noon Pacific time on Saturday.

Subscribers can use those tools through discounted usage bundles or an API key, and get a one-time credit equal to their monthly plan cost. Anthropic’s Boris Cherny said subscriptions “weren’t built for the usage patterns of these third-party tools”.

• Claude subscriptions will no longer cover third-party tools · Boris Cherny on X

• Subscriptions weren’t built for these third-party tools · Boris Cherny on X

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

• Anthropic confirmed testing a new model after a leak named it Claude Mythos.

• A judge blocked part of the Pentagon’s supply chain risk label on Anthropic.

• OpenAI is shutting Sora, and Disney walked away from a planned $1bn stake.

• Arm began selling its own chip, the AGI CPU, with Meta as its lead customer.

• OpenAI won $10bn more in commitments, taking its record round past $120bn.

Browse by day:

• Weekend: Musk Terafab

• Monday: OpenAI PE terms · IWF · Epoch maths

• Tuesday: Sora · OpenAI $10bn · Arm · TurboQuant · Baltimore · LiteLLM

• Wednesday: Manus · GPU smuggling · ARC-AGI-3 · GitHub

• Thursday: Mythos · Anthropic ruling · EU AI Act · US Army · Shield AI · Minab

• Friday: SoftBank facility

Weekend (21-22 March)

Elon Musk launched Terafab, a joint venture of Tesla, SpaceX and xAI to build a chip fab in Austin, Texas.

Musk presented the plan on Saturday in a livestream on X. He said current chipmaking capacity covers about 2% of what Tesla and SpaceX will need: “We either build the Terafab or we don’t have the chips.” Tesla is targeting a 2-nanometre process at the North Campus of Giga Texas.

• TERAFAB Will Build More Compute Than Entire World (Supercut) · Tesla Jigsaw

• Tesla and SpaceX announce $25B ‘Terafab’ chip factory · Electrek

• Elon Musk announces Terafab project · Engadget

Monday 23 March

OpenAI is offering private equity firms a guaranteed minimum return of 17.5% to join enterprise joint ventures, Reuters reported.

Participating firms would also get early access to new OpenAI models. Anthropic is pursuing a similar venture without a comparable guarantee.

• OpenAI sweetens private equity pitch amid turf war with Anthropic · Reuters

The UK’s Internet Watch Foundation (IWF) assessed 8,029 AI-generated images and videos as realistic child sexual abuse material in 2025.

It identified 3,443 AI-generated abuse videos, against 13 in 2024, and classed 65% of the videos as Category A, the most extreme classification.

• AI CSAM Report 2026: Harm Without Limits · IWF

Epoch AI said GPT-5.4 Pro (OpenAI) solved an open Ramsey-style problem on hypergraphs from its FrontierMath: Open Problems set.

Kevin Barreto and Liam Price elicited the solution, and problem contributor Will Brian confirmed it. Brian will write it up for publication.

• A Ramsey-style Problem on Hypergraphs · Epoch AI

Tuesday 24 March

OpenAI is shutting down its Sora video app, and Disney walked away from a planned $1bn stake.

The Sora team announced on X on Tuesday afternoon that it is “saying goodbye to Sora”. Timelines for the app and the API will follow. Chief financial officer Sarah Friar told CNBC that the company is “facing a lack of compute”. Disney ended its OpenAI partnership. The three-year licence would have let Sora use >200 Disney, Marvel, Pixar and Star Wars characters. Disney said it respects “OpenAI’s decision to exit the video generation business”.

• We’re saying goodbye to Sora · Sora on X

• OpenAI Shuts Down Sora AI Video, Disney Drops Planned $1B Investment · Variety

OpenAI secured another $10bn in commitments, taking its record round past $120bn.

Friar said Andreessen Horowitz, D.E. Shaw Ventures, MGX, TPG, T. Rowe Price and Microsoft are in the new tranche. The round’s initial target was $100bn. Asked about an IPO, Friar said OpenAI is “starting to build that outcome”.

• OpenAI secures an extra $10 billion in record funding round, CFO Friar says · CNBC

Arm began selling its own chip, the AGI CPU, with Meta as lead customer.

Arm says it is delivering its own silicon products for the first time in its >35-year history. It claims >2x the performance per rack of the latest x86 systems. Meta co-developed the chip and will build multiple generations of CPUs with Arm. Other launch partners include OpenAI, Cerebras, Cloudflare and SK Telecom.

• Announcing Arm AGI CPU · Arm Newsroom

• Meta Partners With Arm to Develop New Class of Data Center Silicon · Meta

Google Research published TurboQuant, which it says compresses the key-value cache of large language models to 3 bits with no loss of accuracy in its tests.

Google reports at least a 6x cut in key-value memory. At 4 bits, attention runs up to 8x faster on Nvidia H100 GPUs than with unquantised 32-bit keys.

• TurboQuant: Redefining AI efficiency with extreme compression · Google Research

Baltimore sued X, xAI and SpaceX, alleging that Grok’s sexualised deepfakes breach the city’s consumer protection law.

The complaint, filed in the Circuit Court for Baltimore City, says the companies marketed Grok as a safe assistant without disclosing that it could generate explicit deepfakes. The city seeks civil penalties, an injunction, restitution and disgorgement.

• City of Baltimore Sues Over Grok AI’s Role in Generating Deepfakes · DiCello Levitt

Two releases of LiteLLM, an open-source gateway for calling AI model APIs, were published to PyPI with an apparent credential stealer inside.

Versions 1.82.7 and 1.82.8 were live for about 40 minutes on Tuesday before PyPI quarantined them. The maintainers believe the attack came through the Trivy scanner in their build pipeline. Users of the LiteLLM Proxy Docker image were not affected.

• Security Update: Suspected Supply Chain Incident · LiteLLM

Wednesday 25 March

China barred Manus co-founders Xiao Hong and Ji Yichao from leaving the country while it reviews Meta’s $2bn purchase, the FT reported.

Beijing is examining whether the sale broke laws on technology exports and outbound investment. Meta had moved >100 Manus staff into its Singapore office in early March. Meta said the deal “complied fully with applicable law”.

• China reviews $2bn Manus sale as founders barred from leaving · Financial Times

• Beijing’s intervention on Meta’s Manus renews debates on ‘China shedding’ · CNBC

The US Justice Department charged three people with conspiring to smuggle Nvidia AI chips to China through companies in Thailand.

Prosecutors say Tommy Shad English, posing as a Thai buyer, ordered 750 Supermicro servers for about $170m in 2023, 600 of them with export-controlled chips. Stanley Yi Zheng of Hong Kong and Matthew Kelly are also charged. It follows last week’s charges against a Supermicro co-founder.

• Three charged with smuggling AI technology to China · US Justice Department

• Three more charged with trying to smuggle GPUs to China · The Register

The ARC Prize Foundation launched ARC-AGI-3, an interactive benchmark of game-like environments, with $2m in prizes.

Humans with no instructions solved every environment. On the leaderboard, which tests models via API without custom scaffolding, every frontier model scored <1%.

• ARC Prize 2026 · ARC Prize

• ARC-AGI-3 technical report · ARC Prize

GitHub will train its AI models on Copilot interaction data from Free, Pro and Pro+ users from 24 April unless they opt out.

The data covers inputs, outputs, code snippets and associated context. Copilot Business and Copilot Enterprise users are not affected.

• Updates to GitHub Copilot interaction data usage policy · GitHub

Thursday 26 March

Anthropic confirmed it is testing a new model, which a leaked draft calls Claude Mythos.

A draft blog post, left in a publicly searchable data store, named the model and called it “by far the most powerful AI model we’ve ever developed”. Anthropic told Fortune the model is “a step change” and is being trialled by early access customers. The draft describes a new tier of model larger than Anthropic’s Opus line. It says the model is “currently far ahead of any other AI model in cyber capabilities”. Anthropic blamed “human error” in configuring its content management system.

• Anthropic acknowledges new AI model with ‘step change’ in capabilities · Fortune

• Anthropic left details of unreleased model unsecured in security lapse · Fortune

A federal judge blocked part of the Pentagon’s move to label Anthropic a supply chain risk.

Judge Rita Lin wrote that nothing in the statute supports “the Orwellian notion” that an American company may be branded a potential adversary for disagreeing with the government. She called the government’s conduct “classic illegal First Amendment retaliation” and delayed her order by a week to allow an appeal. At Tuesday’s hearing Lin said the Pentagon’s moves look “like an attempt to cripple Anthropic”. Anthropic must still persuade the DC Circuit court.

• Order granting injunction · CourtListener · Hearing transcript · CourtListener

• ‘Premature’: Anthropic still in trouble despite court win · Politico

The European Parliament adopted its position to delay the EU AI Act’s high-risk rules and ban AI nudifier apps.

The vote was 569 to 45, with 23 abstentions. MEPs propose 2 December 2027 for high-risk systems listed in the Act and 2 August 2028 for those covered by EU sectoral law. Negotiations with the Council can now begin.

• AI Act: delayed application, ban on nudifier apps · European Parliament

The US Army conditionally selected Carlyle and CyrusOne, owned by KKR and BlackRock funds, to negotiate building commercial data centres on two bases.

Carlyle would build at Fort Bliss, Texas, and CyrusOne at Dugway Proving Ground, Utah, under an Army programme that leases land on its bases. The sites are due to reach initial operating capability in FY2027 and FY2029. The deal is not yet final.

• Army reaches conditional agreement for hyperscaled data centers · US Army

Defence technology company Shield AI is raising $2bn at a $12.7bn valuation and plans to buy simulation software maker Aechelon Technology.

The package is a $1.5bn Series G led by Advent International plus $500m of fixed-return preferred equity from Blackstone funds. Aechelon’s software supports the Pentagon’s Joint Simulation Environment.

• Shield AI to acquire Aechelon and raise $2B at $12.7B valuation · Shield AI

A Guardian long read argued that the Maven targeting system, not Anthropic’s Claude, lay behind the US strike on a school in Minab, Iran.

The strike on the Shajareh Tayyebeh school on 28 February killed 175-180 people, mostly girls. Kevin T Baker writes that coverage centred on whether Claude chose the target, while “the targeting for Operation Epic Fury ran on a system called Maven”.

• AI got blame for Iran school bombing. The truth is far more worrying · The Guardian

Friday 27 March

SoftBank signed a $40bn bridge facility, mainly to fund its $30bn follow-on investment in OpenAI.

The lenders are JPMorgan Chase, Goldman Sachs, Mizuho Bank, Sumitomo Mitsui and MUFG Bank. The facility carries no collateral and matures on 25 March 2027.

• Bridge Facility Agreement for the Follow-on Investments in OpenAI · SoftBank

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

• Nvidia forecast at least $1tn of Blackwell and Rubin revenue from 2025 to 2027.

• A Supermicro co-founder was charged over AI servers smuggled to China.

• A Meta employee exposed user data to staff after following an AI agent’s advice.

• OpenAI agreed to buy Astral, maker of the uv and Ruff Python tools, for Codex.

• The UK dropped plans to let AI train on copyrighted work unless creators opt out.

Browse by day:

• Weekend: ByteDance pauses Seedance 2.0 launch

• Monday: Nebius · Britannica · xAI · Nvidia GTC · Nscale · Comet

• Tuesday: UK £2.5bn · OpenAI-AWS · Copilot · GPT-5.4 mini · H200 · Anthropic

• Wednesday: UK copyright · Micron · Meta agent · Xiaomi MiMo

• Thursday: Uber-Rivian · Astral · Cursor Composer 2 · Supermicro

• Friday: White House AI framework

Weekend (14-15 March)

ByteDance suspended the global launch of its Seedance 2.0 video model after copyright disputes with Hollywood, The Information reported on Saturday.

ByteDance planned a mid-March release through its BytePlus cloud and a standalone app outside China. There is no new date. Disney, Netflix, Warner Bros. Discovery, Paramount Skydance and Sony had sent cease-and-desist letters.

• ByteDance suspends video AI model after copyright disputes · The Information

Monday 16 March

Meta signed an AI infrastructure deal with Nebius worth up to about $27bn over five years.

Amsterdam-based Nebius will provide $12bn of dedicated capacity on one of the first large Nvidia Vera Rubin deployments from early 2027. Meta will also buy up to $15bn of further capacity that Nebius does not sell to other customers. On Wednesday Nebius priced $4.0bn of convertible notes, upsized from $3.75bn.

• Nebius signs new AI infrastructure agreement with Meta · Nebius

• Nebius announces pricing of upsized $4.0 billion convertible notes · Nebius

Encyclopaedia Britannica and its Merriam-Webster subsidiary sued OpenAI, alleging it scraped nearly 100,000 Britannica articles to train its models.

The complaint also targets ChatGPT outputs containing “full or partial verbatim reproductions” and ChatGPT’s use of Britannica articles in retrieval. A Lanham Act claim says ChatGPT attributes invented information to the publisher.

• Complaint, Encyclopaedia Britannica v. OpenAI · CourtListener

• The dictionary sues OpenAI · TechCrunch

Three Tennessee teenagers filed a class action against xAI, alleging its technology was used to make sexually explicit images of them.

The complaint says the perpetrator used a third-party app that licenses xAI’s Grok model, with the images generated on xAI’s servers. The plaintiffs seek class-action status, damages and an injunction.

• Complaint, Doe 1 v. X.AI Corp. · CourtListener

• Teens sue Musk’s xAI over AI-generated nonconsensual nudes · NPR

Nvidia said it sees at least $1tn of Blackwell and Rubin revenue from 2025 to 2027.

Jensen Huang told the GTC keynote in San Jose that he now sees “at least $1 trillion” of revenue from 2025 to 2027. A year earlier he had cited about $500bn of high-confidence purchase orders for Blackwell and Rubin through 2026. CFO Colette Kress confirmed the figure covers only Blackwell and Rubin plus networking. It excludes Groq LPUs and standalone CPUs.

At GTC Nvidia also launched the Groq 3 LPX inference rack and the Vera CPU for agentic AI. It formed the Nemotron Coalition with labs including Mistral AI, Cursor and Perplexity to build open frontier models.

• NVIDIA GTC 2026: Live Updates on What’s Next in AI · Nvidia

• Jensen Huang says company has one trillion dollars in orders through 2027 · DCD

• NVIDIA Vera Rubin Opens Agentic AI Frontier · Nvidia

• NVIDIA Launches Nemotron Coalition of Leading Global AI Labs · Nvidia

UK AI infrastructure firm Nscale agreed to buy American Intelligence & Power and its West Virginia campus, which it plans to grow to 8 GW by 2031.

The Monarch Compute Campus covers up to 2,250 acres in Mason County, with 2 GW of initial power due by H1 2028. Nscale also signed a letter of intent to provide Microsoft with 1.35 GW of Vera Rubin compute in tranches from late 2027, and will use Caterpillar gas generators for power.

• Nscale Acquires American Intelligence & Power Corporation · Nscale

• Nscale and Microsoft announce 1.35GW of Vera Rubin GPUs · PR Newswire

A US appeals court suspended the injunction barring Perplexity’s Comet agent from shopping in password-protected Amazon accounts.

The Ninth Circuit granted a temporary stay on Monday. On Wednesday it stayed the injunction pending appeal, expedited the case and set it for hearing in June.

• Orders of 16 and 18 March, Amazon v. Perplexity, Ninth Circuit · CourtListener

Tuesday 17 March

Chancellor Rachel Reeves announced a record £2.5bn for UK AI and quantum computing, including up to £1bn to buy quantum computers.

The Treasury said the UK will “achieve the fastest AI adoption in the G7”. Quantum takes £2bn of the package.

• UK will win AI race as Chancellor sets out economic ‘big choices’ · GOV.UK

OpenAI signed a deal with AWS to sell its AI products to US government agencies for classified and unclassified work.

OpenAI said it covers Amazon Bedrock in AWS GovCloud and in AWS Classified Regions for Secret and Top Secret workloads. On Wednesday the Financial Times reported that Microsoft is weighing legal action against Amazon and OpenAI over whether AWS can offer OpenAI’s Frontier without breaching its Microsoft deal. Microsoft said it is “confident that OpenAI understands and respects the importance of living up to [its] legal obligation”.

• OpenAI expands government footprint with AWS deal, report says · TechCrunch

• Microsoft weighs legal action over Amazon-OpenAI cloud deal · Financial Times

Microsoft put former Snap executive Jacob Andreou in charge of a unified Copilot, freeing Mustafa Suleyman to focus on frontier models.

Satya Nadella named Andreou EVP, Copilot, in a memo. Suleyman said he will focus “all my energy on our Superintelligence efforts” and deliver models within five years.

• Announcing Copilot leadership update · Microsoft

OpenAI released GPT-5.4 mini and GPT-5.4 nano, smaller versions of GPT-5.4 for coding, subagents and computer use.

OpenAI says GPT-5.4 mini runs >2x faster than GPT-5 mini and approaches GPT-5.4 on SWE-Bench Pro and OSWorld-Verified. It costs $0.75/m input and $4.5/m output tokens, against $0.25/m and $2/m for GPT-5 mini. API-only GPT-5.4 nano costs $0.20/m and $1.3/m, against $0.05/m and $0.40/m for GPT-5 nano.

• Introducing GPT-5.4 mini and nano · OpenAI

• Introducing GPT-5 for developers · OpenAI

Jensen Huang said Nvidia is restarting manufacturing of its H200 chips for China after receiving purchase orders from “many customers”.

“We’ve been licensed for many customers in China for H200,” Huang told reporters at GTC. Late on Tuesday Reuters reported that Beijing had approved H200 purchases by several Chinese companies.t

• Nvidia restarting production for H200 chips for sales in China · Axios

• Chinese authorities approve Nvidia’s H200 AI chip sales, source says · Reuters

The Justice Department told the court in Anthropic’s lawsuit that the company could disable Claude or alter it during warfighting operations.

The Tuesday filing argues that the supply chain risk designation did not breach Anthropic’s First Amendment rights. A declaration by Pentagon official Emil Michael says Anthropic employs foreign nationals “including many from the [People’s] Republic of China”. Judge Rita Lin set a hearing for the following Tuesday.

In a Friday filing Anthropic said it “does not maintain any back door or remote ‘kill switch’”. Thiyagu Ramasamy, its head of public sector, wrote that it has never been able to stop Claude working or alter its functionality during military operations.

• Opposition to motion · CourtListener · Declaration of Emil Michael · CourtListener

• Declaration of Thiyagu Ramasamy · CourtListener

• DOJ says Anthropic can’t be trusted with warfighting systems · WIRED

Wednesday 18 March

The UK government dropped its plan for an AI copyright exception with opt-out.

The opt-out was the government’s preferred option in its consultation, and most respondents rejected it. A report published on Wednesday says a broad exception with opt-out “is no longer the government’s preferred way forward”. The government will not change copyright law until it is confident reforms meet its aims. It will gather evidence and will not intervene in the licensing market for now.

• Report on Copyright and Artificial Intelligence · GOV.UK

Micron reported fiscal Q2 revenue of $23.9bn, against $8.1bn a year earlier.

Micron guided fiscal Q3 revenue to $33.5bn, plus or minus $750m.

• Micron fiscal Q2 2026 results, Exhibit 99.1 · SEC

A Meta employee exposed sensitive data to unauthorised staff after acting on an AI agent’s advice.

An internal agent posted an answer to a colleague’s question on a company forum without approval. An employee acted on its inaccurate advice and triggered a “SEV1” incident, Meta’s second-highest severity rating. For almost two hours, employees could see company and user data they were not authorised to view. Meta said “no user data was mishandled”.

• Inside Meta, a Rogue AI Agent Triggers Security Alert · The Information

• A rogue AI led to a serious security incident at Meta · The Verge

Xiaomi released MiMo-V2-Pro, a model with >1T parameters, and confirmed it was the anonymous model being tested on OpenRouter.

MiMo-V2-Pro has 42B active parameters and a 1M-token context, and is API-only, with no weights released. Artificial Analysis scored it 49 on its Intelligence Index, behind GLM-5 (Zhipu AI) at 50 and ahead of Kimi K2.5 (Moonshot AI) at 47.

• MiMo-V2-Pro · Xiaomi

• MiMo-V2-Pro: Everything you need to know · Artificial Analysis

Thursday 19 March

Uber will invest up to $1.25bn in Rivian for robotaxis through 2031, with $300m on signing, subject to regulatory approval.

The rest is tied to autonomy milestones. The first phase puts 10,000 autonomous Rivian R2 vehicles on Uber, starting in San Francisco and Miami in 2028. On Monday Uber and Nvidia said they plan robotaxis running entirely on Nvidia software, starting in Los Angeles and San Francisco in H1 2027 and reaching 28 cities by 2028.

• Uber and Rivian partner to deploy up to 50,000 robotaxis · Business Wire

• NVIDIA to Launch L4 Robotaxis on Uber Across 28 Cities by 2028 · Uber

OpenAI agreed to acquire Astral, maker of the Python tools uv, Ruff and ty.

The Astral team will join Codex. OpenAI says Codex has >2m weekly active users after 3x user growth since the start of the year. Founder Charlie Marsh said OpenAI will keep supporting Astral’s open source tools after the deal closes. The tools are downloaded hundreds of millions of times a month.

• OpenAI to acquire Astral · OpenAI

• Astral to join OpenAI · Astral

Cursor released Composer 2, a coding model priced at $0.5/m input and $2.5/m output tokens.

Cursor says the model comes from its first continued pretraining run followed by reinforcement learning (RL). A faster variant costs $1.5/m and $7.5/m. A developer then posted a Cursor model identifier pointing to Kimi K2.5 (Moonshot AI) with RL. On Friday Moonshot AI said Kimi K2.5 is Composer 2’s foundation, accessed through Fireworks AI under an authorised commercial partnership.

• Introducing Composer 2 · Cursor

• A developer’s post · fynn on X · Congrats Cursor on launch Composer 2 · Kimi on X

US prosecutors charged Supermicro co-founder Yih-Shyan Liaw with conspiring to divert AI servers to China.

Prosecutors allege that a Southeast Asian company bought about $2.5bn of servers in 2024 and 2025 on the defendants’ instructions. At least $510m worth allegedly went to China between late April and mid-May 2025. The defendants allegedly staged thousands of dummy servers to fool inspectors. On Friday Supermicro said Liaw had resigned from its board and named DeAnna Luna acting chief compliance officer. Its shares fell 33% that day.

• Three charged over diverting AI technology to China · US Department of Justice

• Supermicro Form 8-K · SEC · Super Micro co-founder indicted · CNBC

Friday 20 March

The White House published an AI policy framework asking Congress to pre-empt state AI laws.

It wants Congress to override state laws that regulate how models are developed or penalise companies for others’ use of their AI, and not to create a new federal AI regulator. It also covers protections for children and teens online. On Wednesday Senator Marsha Blackburn released a discussion draft of the TRUMP AMERICA AI Act, a federal framework that would replace state AI laws. It folds in the Kids Online Safety Act and the NO FAKES Act, and would sunset Section 230.

• National Policy Framework for AI: Legislative Recommendations · White House

• Blackburn releases draft of national policy framework for AI · CFO Dive

ALETH · OPEN ACCESS RESEARCH

aleth.co

The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

• Anthropic sued the Pentagon to overturn its blacklisting, and Microsoft backed it.

• Nvidia released Nemotron 3 Super and plans to spend $26bn on open models.

• Meta delayed its next model to at least May after trailing rivals.

• Yann LeCun’s AMI Labs raised $1.0bn for world models at $3.5bn pre-money.

• OpenAI robotics lead Caitlin Kalinowski resigned over the Pentagon deal.

Browse by day:

• Weekend: Kalinowski quits OpenAI

• Monday: UK AI fund · Nscale · UK pledges · Anthropic sues · McKinsey hack

• Tuesday: AMI Labs · Thinking Machines · Moltbook · Perplexity

• Wednesday: OpenClaw curbs · CCDH · Nebius · Nemotron · Atlassian

• Thursday: Wayve Tokyo · Ukraine · Meta delay · ByteDance · xAI

• Friday: Cerebras · Meta cuts · Chip rule

Weekend (7-8 March)

Caitlin Kalinowski resigned from OpenAI over its Pentagon deal.

Kalinowski, who led OpenAI’s hardware and robotics teams from November 2024, said surveillance of Americans without judicial oversight and lethal autonomy without human authorisation “are lines that deserved more deliberation than they got”.

OpenAI said its agreement makes clear its red lines: “no domestic surveillance and no autonomous weapons”.

• I resigned from OpenAI · Caitlin Kalinowski on X

• OpenAI robotics leader resigns over surveillance and autonomous weapons · Fortune

Monday 9 March

The UK government unveiled Sovereign AI, a £500m venture fund to back British AI founders “from day one”.

It combines cash with access to the UK’s largest AI computing capacity, unique data sets and government procurement routes. The fund opens on 16 April.

The government first announced the unit in November, with venture capitalist James Wise as chair, to help “build and scale AI capabilities on British shores”.

• Introducing Sovereign AI, the Government’s new £500m fund · Sovereign AI on X

• The UK’s £500 million fund for AI, launching 16 April · Sovereign AI

• AI to power national renewal as government announces billions · GOV.UK

UK AI infrastructure company Nscale raised a $2bn Series C at a $14.6bn valuation and named Sheryl Sandberg, Nick Clegg and Susan Decker to its board.

Aker and 8090 Industries co-led the round. Its total includes a $433m pre-Series C SAFE (simple agreement for future equity) from October.

• Nscale raises $2 billion in Series C, the largest in European history · Nscale

• Sandberg, Clegg join Nscale board as it hits $14.6B valuation · TechCrunch

A Guardian investigation found the UK government is not auditing the AI investment pledges from Nscale and CoreWeave that it has promoted.

CoreWeave’s £1bn pledge, billed in 2024 as bringing “two new data centres to our shores”, instead made it a customer of two existing data centres, built in 2002 and 2015. The government said Nscale’s $2.5bn UK investment was “not a formal contract, rather an intention to commit capital”.

• Revealed: UK’s multibillion AI drive is built on ‘phantom investments’ · The Guardian

Anthropic sued the Pentagon to overturn its supply chain risk designation.

Anthropic filed a complaint in California against 17 federal agencies and the Executive Office of the President, with a motion for a temporary restraining order. A separate petition asks the US Court of Appeals to review the designation. The complaint says the Department of War threatened two contradictory moves: invoking the Defense Production Act to force Anthropic to hand over Claude, or banning it as a security risk. It argues the statute used, 10 USC 3252, was written for foreign adversaries sabotage of information systems.

Chief financial officer Krishna Rao said in a filing that hundreds of millions of dollars of expected Pentagon-related revenue this year is at risk, and that wider pressure could cost billions in sales. Anthropic’s all-time sales exceed $5bn. Hours after the filing, over 30 OpenAI and Google employees, including Google DeepMind chief scientist Jeff Dean, filed an amicus brief in its support. Microsoft followed on Tuesday, asking for a restraining order “for all existing contracts”. Without one, it said, contractors must “act immediately to alter existing product and contract configurations”.

• Complaint for relief · CourtListener · Motion for restraining order · CourtListener

• Petition for review, DC Circuit · CGPTiEW · Declaration Krishna Rao · CourtListener

• Amici curiae: OpenAI/Google · CourtListener ; Microsoft, Anthropic · CourtListener

• Microsoft backs Anthropic in Pentagon battle, urges restraining order · CNBC

Security start-up CodeWall said its AI agent gained full read and write access to the database of Lilli, McKinsey’s internal AI platform, within two hours.

The agent found a SQL injection flaw without credentials, exposing 46.5 million chat messages stored in plaintext. CodeWall said it alerted McKinsey on 1 March and the firm patched the exposed endpoints a day later.

• How we hacked McKinsey’s AI platform · CodeWall

Tuesday 10 March

Yann LeCun’s AMI Labs raised $1.0bn to build world models.

The round values the Paris-based lab at $3.5bn before the new money. Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions co-led it, and Nvidia, Samsung, Temasek and Toyota Ventures are among its backers.

Chief executive Alexandre LeBrun said AMI has no plans to generate revenue for now.

• Unveiling our new startup Advanced Machine Intelligence · Yann LeCun on X

• Yann LeCun’s AMI Labs raises $1.03B to build world models · TechCrunch

Thinking Machines Lab agreed with Nvidia to deploy at least 1 GW of Vera Rubin systems for its frontier model training.

Deployment is targeted for early next year. Nvidia has also made a “significant investment” in the lab, which Mira Murati co-founded and runs.

• NVIDIA and Thinking Machines Lab announce gigawatt-scale partnership · Nvidia

• Thinking Machines and NVIDIA announce gigawatt partnership · Thinking Machines

Meta agreed to acquire Moltbook, a social network built for AI agents, and will bring its creators Matt Schlicht and Ben Parr into Meta Superintelligence Labs.

• Facebook parent Meta acquires Moltbook, an AI agent social network · Axios

A US judge granted Amazon an injunction on Monday blocking Perplexity’s Comet agent from shopping in password-protected Amazon accounts.

US District Judge Maxine Chesney found Amazon likely to succeed on its Computer Fraud and Abuse Act claims, because Comet acts with the Amazon user’s permission but without Amazon’s authorisation. The order is stayed for seven days for an appeal.

• Order granting Amazon’s motion for preliminary injunctive relief · CourtListener

Wednesday 11 March

Chinese state agencies and companies, including the largest banks, were warned against installing OpenClaw on office devices, Bloomberg reported.

Some staff were barred from installing it on office computers and on personal phones using company networks. Other notices required only prior approval.

• China curbs state use of OpenClaw AI just as Tencent embraces it · News24

Eight in 10 chatbots tested by the Center for Countering Digital Hate (CCDH) and CNN were willing to help users posing as teenagers plan violent attacks.

The ten included ChatGPT (OpenAI), Google Gemini and Microsoft Copilot. CCDH said only Claude (Anthropic) and Snapchat’s My AI consistently refused, Claude in 68% of cases. Perplexity and Meta AI helped in 100% and 97% of responses.

• Killer Apps · CCDH

Nvidia will invest $2bn in Nebius under a partnership to help the AI cloud provider deploy >5 GW of Nvidia systems by the end of 2030.

Nebius will be an early adopter of Nvidia’s Rubin platform and Vera CPUs.

• NVIDIA and Nebius partner to scale full-stack AI cloud, Exhibit 99.1 · SEC

Nvidia released Nemotron 3 Super and plans to spend $26bn on open models.

Nemotron 3 Super is a 120-billion-parameter hybrid mixture of experts (MoE) model, with 12 billion of its parameters active at inference, released with open weights under a permissive licence. Nvidia says it delivers up to 5x higher throughput than the previous Nemotron Super model.

Nvidia will spend $26bn over the next five years building open models, according to a 2025 financial filing that WIRED reported and Nvidia executives confirmed.

• New NVIDIA Nemotron 3 Super delivers 5x higher throughput for agentic AI · Nvidia

• Nvidia will spend $26 billion to build open-weight AI models, filings show · WIRED

Atlassian is cutting about 1,600 jobs, 10% of its workforce, to fund investment in AI and enterprise sales.

Chief executive Mike Cannon-Brookes said the cuts would “self-fund further investment in AI and enterprise sales”, and that AI is not replacing Atlassian’s employees. A filing puts the charges at $225m to $236m.

• An important update on our team · Atlassian

• Atlassian slashes 10% of workforce to ‘self-fund’ investments in AI · CNBC

Thursday 12 March

Uber, Wayve and Nissan plan to launch a robotaxi service in Tokyo, with a pilot scheduled for late 2026.

UK-based Wayve will put its self-driving software into a Nissan Leaf offered on Uber’s network. It is Uber’s first robotaxi partnership in Japan.

• Wayve, Uber and Nissan announce collaboration on robotaxis · Nissan

Ukraine’s government approved a framework letting partners and Ukrainian companies train AI models for unmanned systems on its battlefield data.

Defence minister Mykhailo Fedorov said the dataset includes “millions of annotated frames collected during tens of thousands of combat flights”. A Defence Ministry platform lets models be trained without direct access to sensitive databases.

• Ukraine first in the world to open AI training on real battlefield data · ArmyInform

Meta delayed its next AI model to at least May.

The model was due in mid-March but fell short of leading models from Google, OpenAI and Anthropic on reasoning, programming and writing in internal tests, The New York Times reported. Meta’s leadership discussed temporarily licensing Google’s Gemini, but made no decision. Meta says updates are coming “very soon”.

• Meta delays new AI model after performance concerns · The New York Times

ByteDance is working with Aolani Cloud on about 500 Nvidia Blackwell systems in Malaysia, roughly 36,000 B200 chips, The Wall Street Journal reported.

The hardware would likely cost >$2.5bn. US export controls prevent ByteDance getting direct access to the chips in China.

• China’s ByteDance gets access to top Nvidia AI chips · The Wall Street Journal

Elon Musk said xAI “was not built right first time around, so is being rebuilt from the foundations up”, after co-founders Zihang Dai and Guodong Zhang left.

The two left after Musk complained that xAI’s coding tools were not competing with Claude Code (Anthropic) or Codex (OpenAI). Manuel Kroiss and Ross Nordeen are the only co-founders still there. The Financial Times reported that SpaceX and Tesla executives were sent in to evaluate staff.

• xAI was not built right first time around · Elon Musk on X

• ‘Not built right the first time’: Musk’s xAI is starting over again, again · TechCrunch

Friday 13 March

AWS and Cerebras will offer AI inference on Amazon Bedrock that splits each job between AWS Trainium servers and Cerebras CS-3 systems.

Trainium handles prefill and the CS-3 handles decode. The service launches “in the next couple of months”, and AWS later plans to offer open-source large language models and Amazon Nova on Cerebras hardware.

• AWS and Cerebras aim to set a new standard for AI inference speed · Amazon

Meta is planning layoffs that could reach 20% or more of its workforce as AI infrastructure costs mount, Reuters reported.

No date has been set and the size is not final, three sources said. Meta called it “speculative reporting about theoretical approaches”.

• Exclusive: Meta planning sweeping layoffs as AI costs mount · Reuters

The US Commerce Department withdrew a planned rule that would have tightened export controls on AI chips, Reuters reported.

The draft, “AI Action Plan Implementation”, had been under White House review since 26 February. A US official said it “was always a draft and remains a draft”.

• AI Action Plan Implementation, RIN 0694-AJ90: withdrawn 13 March · OIRA

• US Commerce Department withdraws planned rule on AI chip exports · Reuters

ALETH · OPEN ACCESS RESEARCH

aleth.co