ALETH / AI-BRIEF / 2026-04-11 / MYTHOS TOO GOOD AT HACKING FOR PUBLIC RELEASEThe 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