ALETH / AI-BRIEF / 2026-05-16 / AI CHIPMAKER CEREBRAS SOARS IN $5.6BN DEBUTThe Aleth Briefs link to the sources behind the stories and show how the week unfolded.
The week in five lines:
Cerebras priced its IPO at $185, raising $5.6bn, and rose 68% on debut.
Google caught criminals using a zero-day it believes AI built, a first.
Mythos Preview became the first model to complete both UK AISI cyber ranges.
Isomorphic Labs raised $2.1bn, with the UK Sovereign AI Fund among the backers.
The US cleared ten Chinese firms to buy Nvidia H200s, but none has shipped.
Browse by day:
Weekend: Baidu ERNIE 5.1 · FSU lawsuit
Monday: Google · DeployCo · Microsoft · Musk trial · TanStack · Thinking Machines
Tuesday: Isomorphic · Microsoft MDASH
Wednesday: UK AISI · Cerebras · Recursive · Fractile · OpenAI-Illinois
Thursday: US H200 · Calif · Anthropic-Gates · HMRC-Quantexa
Friday: OpenAI reorg · ChatGPT finance
Weekend 9-10 May
Baidu released ERNIE 5.1 and said it cost about 6% of what comparable models need to pre-train.
It has a third of ERNIE 5.0’s total parameters and half its active ones, and builds on ERNIE 5.0’s pre-training.
ERNIE 5.1 officially released! · Baidu
The widow of a victim of last year’s Florida State University shooting sued OpenAI, alleging ChatGPT helped the gunman plan the attack.
The complaint says the alleged shooter spent months discussing guns and mass shootings with ChatGPT. It says the chatbot told him a school shooting usually needs “3 or more dead” to draw national attention. OpenAI told NBC News that ChatGPT gave only generally available information. Florida’s attorney general opened a criminal investigation into OpenAI in late April.
Joshi v OpenAI Foundation, complaint · CourtListener
OpenAI sued over ChatGPT’s alleged role guiding FSU shooter · NBC News
Monday 11 May
Google said criminals used a zero-day exploit it believes was built with AI, the first such case it has seen.
Google Threat Intelligence Group (GTIG) said the attackers planned a mass exploitation event, which its discovery may have prevented. A zero-day is a flaw unknown to the software’s maker, so no fix exists when it is exploited. The exploit bypassed two-factor authentication on a popular open-source web administration tool but still needed valid user credentials. Google disclosed the flaw to the vendor.
Adversaries leverage AI for vulnerability exploitation · Google Cloud
OpenAI formally launched the OpenAI Deployment Company and agreed to buy consultancy Tomoro to staff it.
Tomoro would bring about 150 forward deployed engineers, and its clients include Tesco and Virgin Atlantic. OpenAI majority-owns and controls the new company, which launches with >$4bn of initial investment and 19 partner firms led by TPG.
OpenAI launches DeployCo to help businesses build around intelligence · OpenAI
OpenAI’s revenue-share payments to Microsoft will be capped at $38bn, The Information reported.
The Information put OpenAI’s saving through 2030 at $97bn. Microsoft had said in April that payments would continue at the previously agreed percentage, subject to an overall cap. Neither company responded to Reuters.
OpenAI to save $97 billion through 2030 in latest Microsoft deal · The Information
OpenAI, Microsoft agree to cap revenue sharing at $38 billion · Reuters
Satya Nadella, Ilya Sutskever and Sam Altman testified in Musk v Altman before the trial drew to a close.
Nadella said Elon Musk never told him Microsoft’s OpenAI investments broke any special terms, and called the board’s firing of Altman “amateur city”. Sutskever testified that his OpenAI shares are worth about $7bn. Altman testified that Musk once suggested control of OpenAI “should pass to my children”. A filing showed Altman holds >$2bn in companies that work with OpenAI. Closing arguments are done and a verdict is expected next week.
OpenAI trial: Nadella says Musk never asked him about Microsoft deals · CNBC
Altman forced to confront claims that he’s a prolific liar · Ars Technica
Is OpenAI changing its tune on AI laws? · Transformer
Attackers published 84 malicious versions of 42 TanStack npm packages in the Mini Shai-Hulud campaign, which also reached Mistral AI and OpenAI.
The malware steals cloud, GitHub and SSH credentials and spreads through its victims’ other packages. It was spotted within 20-26 minutes. Microsoft found a compromised Mistral AI PyPI package, and Aikido said Mistral’s npm SDKs were hit too. OpenAI said two employee devices were affected and limited credential material was taken. The repositories held its signing certificates, which it is rotating. macOS users must update.
Postmortem: TanStack npm supply-chain compromise · TanStack
Our response to the TanStack npm supply chain attack · OpenAI
Compromised Mistral AI and TanStack packages · Tom’s Hardware
Thinking Machines announced interaction models, its first model, which takes in audio, video and text and responds in real time.
The lab trained the model from scratch to interleave 200 ms slices of input and output instead of waiting for a speaker to finish. It claims the model beats GPT-Realtime-2 (OpenAI) and Gemini Live (Google) on interaction quality. The lab said it would open a limited research preview in the coming months, with a wider release later this year.
Interaction models: a scalable approach · Thinking Machines Lab
Tuesday 12 May
Isomorphic Labs raised $2.1bn in a Series B led by Thrive Capital.
London-based Isomorphic counts the UK Sovereign AI Fund among its new investors alongside MGX, Temasek and CapitalG. Alphabet and GV also invested.
The money will expand its IsoDDE drug design engine, move its drug candidates closer to clinical trials and pay for hiring. It already has partnerships with Novartis, Eli Lilly and Johnson & Johnson.
Isomorphic Labs announces Series B investment round · Isomorphic Labs
Microsoft said its new multi-agent security system helped find 16 Windows vulnerabilities, four of them critical remote code execution flaws.
The system, codenamed MDASH, coordinates >100 specialised AI agents across frontier and distilled models to find, debate and prove exploitable bugs. It found all 21 bugs planted in a private test driver with no false positives.
Defense at AI speed: Microsoft’s new multi-model agentic security system · Microsoft
Wednesday 13 May
Claude Mythos Preview became the first model to complete both of the UK AI Security Institute’s cyber ranges.
The AI Security Institute (AISI) said a newer checkpoint of Anthropic’s model solved The Last Ones in 6 of 10 attempts and the previously unsolved Cooling Tower in 3 of 10. GPT-5.5 (OpenAI) solved The Last Ones in 3 of 10. The ranges simulate attacks on small, undefended corporate networks where the attacker already has a foothold.
In February AISI estimated that frontier models’ 80%-reliability cyber time horizon on its narrow task suite had been doubling every 4.7 months, down from 8 months in its November 2025 estimate. Both models have since run well ahead of that trend.
How fast is autonomous AI cyber capability advancing? · AISI
Cerebras priced its IPO at $185 a share and closed 68% higher on its Nasdaq debut.
Cerebras had lifted its range from $115-125 to $150-160 and the offer to 30m shares. The pricing raised $5.6bn and valued it at $56.4bn fully diluted. Bloomberg had reported that Arm and SoftBank tried to buy it weeks before the listing. The shares opened at $350 on Thursday and closed at $311. By proceeds it was the largest US tech IPO since Uber’s in 2019. The Mohamed bin Zayed University of Artificial Intelligence in the UAE supplied 62% of Cerebras’s 2025 revenue and G42 24%.
Cerebras announces IPO pricing at $185 per share · Cerebras
Cerebras S-1/A #2 · SEC EDGAR
Cerebras prices IPO above expected range · CNBC
Cerebras (CBRS) starts trading on Nasdaq after IPO · CNBC
London-based Recursive Superintelligence came out of stealth with >$650m at a $4.65bn valuation to build AI that improves itself.
GV and Greycroft led the round, with Nvidia and AMD taking part. Chief executive Richard Socher was chief scientist at Salesforce, and co-founder Tim Rocktäschel is a UCL professor and former Google DeepMind scientist. The start-up, which has <30 staff, says it will first focus on “creating AI that improves AI”.
Recursive launch post · Recursive on X
Recursive Superintelligence emerges from stealth with $650M raise · Tech.eu
Fractile raised $220m in a Series B led by Accel, Factorial and Founders Fund.
The London-based company builds inference chips that put compute inside memory, to get round the memory bottleneck that slows GPUs. The money will take its first chips and systems towards customer deployment.
Chips startup Fractile raises $220 million to speed up AI queries · WSJ
Fractile: the future of AI inference · Accel
OpenAI endorsed Illinois SB 315, a frontier AI safety bill that would require third-party audits, along with the US Kids Online Safety Act.
It was the first time OpenAI had used its policy newsletter to endorse legislation. Last month it backed Illinois SB 3444, which would offer AI companies a liability shield. This week OpenAI’s Caitlin Niedermeyer told the Illinois Senate in written testimony that it does not support that safe harbour. Anthropic also endorses SB 315.
Intelligence as a utility · OpenAI Global Affairs
Is OpenAI changing its tune on AI laws? · Transformer
Thursday 14 May
The US cleared about ten Chinese firms, including Alibaba and ByteDance, to buy Nvidia H200 chips, Reuters reported.
Tencent and JD.com were also cleared, and each can buy up to 75,000 chips. None has shipped: Commerce Secretary Howard Lutnick said Beijing is blocking the purchases to protect its own chip industry. Jensen Huang joined President Trump in Beijing, where Trump said export controls came up but nothing concrete followed.
US Treasury Secretary Scott Bessent said the two countries will set up a protocol on AI best practices to keep powerful models from non-state actors.
Exclusive: US clears H200 chip sales to 10 China firms · Reuters
U.S. China AI talks: Bessent says U.S. leads, safety protocol planned · CNBC
Security firm Calif said Mythos Preview helped it build the first public macOS kernel memory corruption exploit to survive Apple’s M5 memory safety system.
Memory Integrity Enforcement is Apple’s hardware-assisted memory safety system, introduced as the marquee security feature of the M5 and A19. Calif found the two bugs on 25 April and had a working exploit by 1 May, taking an unprivileged user on macOS 26.4.1 to a root shell. Calif gave Apple the report and will publish its 55-page write-up once Apple ships a fix.
First public macOS kernel memory corruption exploit on Apple M5 · Calif
Anthropic and the Gates Foundation partnered, committing $200m in grants, Claude usage credits and technical support over four years.
The programmes cover global health, life sciences, education and economic mobility. The largest part goes to health in low- and middle-income countries, including work to speed up development of new vaccines and therapies.
Anthropic partners with the Gates Foundation · Anthropic
HMRC awarded London-based Quantexa a £175m ($236m), 10-year contract to modernise its data and deploy AI.
The programme will rebuild the UK tax authority’s core data infrastructure to help identify tax at risk, and lay the groundwork for advanced AI. Quantexa called it one of the public sector’s largest decision intelligence deployments.
Quantexa selected by HMRC for £175m data and AI transformation · Quantexa
Friday 15 May
OpenAI made Greg Brockman its head of product strategy and folded ChatGPT, Codex and its developer API into one core product team.
Brockman had overseen products while Fidji Simo, chief executive of AGI deployment, was on medical leave. In a memo seen by Wired, he said OpenAI is consolidating its product efforts “to execute with maximum focus toward the agentic future”.
Greg Brockman officially takes control of OpenAI’s products · Wired
OpenAI launched a preview of personal finance in ChatGPT, letting US Pro users connect their bank accounts through Plaid.
Users can link accounts at over 12,000 financial institutions and ask questions about their own spending, saving and debt. ChatGPT cannot see account numbers or change accounts. Plaid said over 200m people a month ask ChatGPT about personal finance.
ChatGPT for personal finance and budgeting · OpenAI
What ChatGPT’s new experience signals for digital finance · Plaid
Aleth · Open Access Research