ALETH / AI-BRIEF / 2026-05-02 / CHINA BLOCKS META’S $2BN MANUS DEALThe Aleth Briefs link to the sources behind the stories and show how the week unfolded.
The week in five lines:
China ordered Meta to unwind its $2bn purchase of the agent start-up Manus.
David Silver’s Ineffable raised $1.1bn in Europe’s largest-ever seed round.
Microsoft ended its exclusive deal, so OpenAI can sell through any cloud.
Musk v Altman went to trial in Oakland over OpenAI’s founding mission.
Google agreed the Pentagon may use its AI for any lawful government purpose.
Browse by day:
Weekend: PocketOS database wipe
Monday: China-Manus · Microsoft-OpenAI · Ineffable · Musk v Altman
Tuesday: Google-Pentagon · Mistral 3.5 · Hua Hong curbs
Wednesday: Robotaxis · Tumbler Ridge · Meta capex · Samsung · Biohub
Thursday: GPT-5.5-Cyber · Five Eyes · Grok 4.3 · PyTorch · GUARD Act
Friday: Pentagon classified AI
Weekend 25-26 April
A Cursor agent running Claude Opus 4.6 deleted PocketOS’s production database and its backups in nine seconds, its founder said.
Decrypt reported that the agent hit a credential mismatch while working in staging, and later produced a written explanation admitting it had broken safety rules. PocketOS founder Jeremy Crane said its most recent recoverable backup was three months old. Railway’s Jake Cooper told Decrypt that Railway recovered the data 30 minutes after he connected with Crane and has since patched the deletion endpoint.
An AI agent just destroyed our production data · Jeremy Crane
AI Agent Deletes Startup’s Database in 9 Seconds, Founder Says · Decrypt
Monday 27 April
China ordered Meta to unwind its $2bn acquisition of Manus.
China’s National Development and Reform Commission (NDRC) said its foreign investment security review office had prohibited the foreign acquisition of Manus and required the parties to withdraw the deal. TechCrunch reported that it gave no explanation and that Meta said the transaction “complied fully with applicable law”.
About 100 Manus employees had already moved into Meta’s Singapore offices by March. On 1 May The Information reported that Moonshot AI, DeepRoute.ai and StepFun are weighing unwinding their foreign holding structures to register in China.
Security review decision on the foreign acquisition of Manus · NDRC
China blocks Meta’s $2B Manus deal after months-long probe · TechCrunch
Chinese firms weigh overhaul after Meta-Manus reversal · The Information
Microsoft and OpenAI rewrote their deal so OpenAI can sell through any cloud.
Microsoft remains OpenAI’s primary cloud partner and keeps a licence to its models and products through 2032, now non-exclusive, with products shipping first on Azure “unless Microsoft cannot and chooses not to support the necessary capabilities”. Microsoft stops paying OpenAI a revenue share, while OpenAI’s payments to it continue through 2030 at the same percentage, subject to a cap.
A day later OpenAI’s models, Codex and managed agents came to Amazon Bedrock in preview.
The next phase of the Microsoft-OpenAI partnership · Microsoft
The next phase of the Microsoft OpenAI partnership · OpenAI
Amazon Bedrock now offers OpenAI models, Codex, and Managed Agents · AWS
OpenAI models, Codex, and Managed Agents come to AWS · OpenAI
David Silver’s London lab Ineffable Intelligence raised a $1.1bn seed round, the largest in European history.
Silver led the DeepMind team that built AlphaGo and AlphaZero. Sequoia Capital co-led the round. TechCrunch put the valuation at $5.1bn. The British Business Bank invested $20m, with the UK government’s Sovereign AI Fund co-investing.
Partnering with Ineffable Intelligence: a superlearner · Sequoia Capital
Sovereign AI invests in Ineffable Intelligence · British Business Bank
UK backs company building AI that can discover new knowledge · GOV.UK
David Silver raised $1.1B for AI that learns without human data · TechCrunch
Musk v Altman went to trial in Oakland over the founding mission of OpenAI.
Judge Yvonne Gonzalez Rogers expects the liability phase to end by 21 May, and Musk’s lawyers said in January he should receive up to $134bn in “wrongful gains”, which he has since asked to be directed back to the OpenAI charity. In opening statements on Tuesday, Musk’s lawyer Steven Molo told jurors that Sam Altman and Greg Brockman “stole a charity”. Emails shown in court had Musk trying to hire OpenAI researchers and halting promised funding in 2017. Asked whether xAI had distilled OpenAI’s technology, Musk answered “Partly”.
Judge in Musk v. Altman seats nine-person jury. Opening arguments next · CNBC
How Elon Musk squeezed OpenAI: they ‘are gonna want to kill me’ · WIRED
Elon Musk faces contentious questions at OpenAI trial · New York Times
How Shivon Zilis operated as Elon Musk’s OpenAI insider · WIRED
Tuesday 28 April
Google agreed to let the Pentagon use its AI for any lawful government purpose.
The Information reported that Google signed a classified deal letting the Pentagon use its models for “any lawful government purpose”. Google said it amends its existing government contract. 9to5Google reported that the agreement requires it to help adjust its AI safety settings and filters at the government’s request. On Monday, >600 Google employees called on the company to reject the deal. Pentagon chief digital and AI officer Cameron Stanley later confirmed expanded use of Gemini, saying “overreliance on one vendor is never a good thing”.
Google and Pentagon reportedly agree on deal for ‘any lawful’ use of AI · The Verge
Pentagon deal uses Gemini for ‘any lawful government purpose’ · 9to5Google
Google employees urge CEO to reject ‘inhumane’ military AI use · Euronews
Pentagon AI chief confirms work with Google after Anthropic blacklist · CNBC
Mistral AI released Mistral Medium 3.5, a multimodal model for agentic and coding work, as open weights under a modified MIT licence.
Its model card says the licence carves out exceptions for companies with large revenue, and that the dense 128B model, with a 256,000-token context window, replaces Medium 3.1 and Magistral in Le Chat and Devstral 2 in Mistral’s Vibe coding agent. Mistral’s own benchmarks give it 77.6% on SWE-Bench Verified.
Mistral Medium 3.5 · Mistral AI
Mistral Medium 3.5 128B model card · Mistral AI on Hugging Face
The US ordered chip-equipment makers to halt some shipments to Hua Hong, China’s second-largest chipmaker, Reuters reported.
Hua Hong’s subsidiary Huali Microelectronics is starting to build a 7-nm production line in Shanghai, Tom’s Hardware reported. The affected US toolmakers include Applied Materials, KLA and Lam Research.
US orders chip equipment firms to halt some shipments to Hua Hong · Reuters
US stops exports of tools to China’s number two chip maker · Tom’s Hardware
Wednesday 29 April
China has suspended new licences for autonomous vehicles after robotaxis stalled in Wuhan, Bloomberg reported.
Dozens of Baidu’s Apollo Go vehicles stopped in the road on the evening of 31 March, with estimates ranging from a few dozen to >200, electrive reported. Nikkei Asia reported that a final decision is not expected before the end of May.
System failure incident prompts China to suspend robotaxi licensing · electrive.com
Seven families of Tumbler Ridge shooting victims sued OpenAI and Sam Altman, alleging the company and ChatGPT were complicit in the injuries or deaths.
The suits, filed in federal court in San Francisco, seek unspecified damages and orders requiring OpenAI to notify law enforcement when its systems flag a risk of real-world harm and to submit to independent monitoring. They allege negligence and product liability.
Complaint, Stacey v. Altman (N.D. Cal.) · CourtListener
Families of Tumbler Ridge shooting victims sue OpenAI and CEO Sam Altman · CNN
Tumbler Ridge, B.C. lawsuits filed in California court against OpenAI · Global News
Meta raised its 2026 capital expenditure guidance to $125-145bn from $115-135bn, citing higher component pricing.
It also cited additional data centre costs for future capacity. Alphabet reported Google Cloud revenue up 63% to $20.0bn in Q1, and Microsoft said its AI business passed an annual revenue run rate of $37bn, up 123% YoY.
Meta reports first quarter 2026 results · Meta
Alphabet announces first quarter 2026 results · Alphabet
FY26 Q3 press release · Microsoft
Samsung Electronics reported record Q1 revenue of KRW 133.9 trillion and operating profit of KRW 57.2 trillion as AI demand lifted its memory business.
Revenue rose 43% QoQ. Nikkei Asia reported that Samsung warned the memory shortage will worsen next year, with customers already placing 2027 orders.
Samsung Electronics announces first quarter 2026 results · Samsung
Samsung warns memory shortage to deepen next year · Nikkei Asia
Biohub launched the Virtual Biology Initiative, a $500m commitment to build an open data foundation for AI models of the human cell.
It will invest in imaging, engineering, data generation and infrastructure that give the scientific community a high-resolution view of the cell.
Virtual Biology Initiative - $500M for AI-powered biology · Biohub
Thursday 30 April
Sam Altman said OpenAI would soon roll out GPT-5.5-Cyber to critical cyber defenders, the restricted access he criticised Anthropic for with Mythos.
TechCrunch noted that Altman had attacked Anthropic for releasing Mythos only to select users. An OpenAI spokesperson said its Trusted Access for Cyber programme has scaled “to thousands of verified defenders and hundreds of teams”.
The same day the UK AI Security Institute found GPT-5.5 is the second model, after Claude Mythos Preview, to solve a multi-step cyber-attack simulation end to end. On Expert-level tasks it averaged a 71.4% pass rate, against 68.6% for Mythos Preview.
Sam Altman on X: starting rollout of GPT-5.5-Cyber · Sam Altman
After dissing Anthropic, OpenAI restricts access to Cyber, too · TechCrunch
Our evaluation of OpenAI’s GPT-5.5 cyber capabilities · UK AI Security Institute
Cyber agencies from the US, UK, Australia, Canada and New Zealand published joint guidance on the careful adoption of agentic AI services.
The US Cybersecurity and Infrastructure Security Agency (CISA) said the autonomy and interconnectedness of agents introduce risks including privilege escalation, emergent behaviours and accountability gaps. The NSA’s release recommends deploying agents incrementally with human oversight, and CyberScoop reported that the guidance warns most organisations give agents far more access than they can safely monitor or control.
Guidance on agentic AI systems · US National Security Agency
CISA and partners release guidance on adopting agentic AI services · CISA
US and allies publish guidance on how to safely deploy AI agents · CyberScoop
xAI released Grok 4.3, which scores 53 on the Artificial Analysis Intelligence Index at about 40% lower input and 60% lower output prices than Grok 4.2.
VentureBeat reported API rates of $1.25/m input and $2.50/m output tokens up to 200,000 input tokens, above which costs double, and a 1 million-token context window. xAI also launched Custom Voices, which clones a voice from a few seconds of audio for its speech APIs behind a two-stage verification check.
xAI launches Grok 4.3 with lower pricing · Artificial Analysis
xAI launches Grok 4.3 at a low price and a voice cloning suite · VentureBeat
Two versions of the PyTorch Lightning package on PyPI, 2.6.2 and 2.6.3, were compromised with malware that steals credentials on import.
Lightning AI’s advisory, published on 30 April, said the affected versions include a credential-harvesting mechanism. Semgrep said the Mini Shai-Hulud themed code also steals tokens and cloud secrets and tries to poison GitHub repositories.
Compromise of PyTorch Lightning PyPI package versions · Lightning AI
Shai-Hulud themed malware found in the PyTorch Lightning library · Semgrep
The Senate Judiciary Committee unanimously advanced the GUARD Act, a bipartisan bill that would bar children from AI companions.
Senators Josh Hawley and Richard Blumenthal co-sponsored the bill, which would require age verification before use and permit fines of up to $100,000 per violation, The Record reported.
Judiciary Committee advances two bipartisan bills · Senate Judiciary Committee
Senate panel advances bill barring minors from AI companions · The Record
Friday 1 May
The Pentagon signed agreements with eight companies to deploy their AI on classified networks, and Anthropic is not among them.
The companies are SpaceX, OpenAI, Google, Nvidia, Reflection AI, Microsoft, Amazon Web Services and Oracle, which will deploy on both IL6 and IL7 environments, DefenseScoop reported. Oracle said its agreement covers “lawful use”. Anthropic remains in litigation with the department.
Classified networks AI agreements · US Department of War
Department of War deal with Oracle to deploy AI on classified networks · Oracle
DOD expands classified AI work, excluding Anthropic · DefenseScoop
ALETH · OPEN ACCESS RESEARCH