Today in AI — 8 April 2026

The labs are building walls. Anthropic won't release its most capable model publicly. Meta is walking back open-source. OpenAI, Anthropic, and Google are sharing intelligence to stop model distillation. A year ago the industry couldn't stop talking about democratisation. Today the conversation is about control.

Security and the model too dangerous to ship

Anthropic's Project Glasswing deploys Claude Mythos exclusively to 12 companies for defensive cybersecurity. The model found thousands of zero-day vulnerabilities across every major OS and browser. When a lab decides its own product is too risky for its own customers, the capability overhang is real.

• Anthropic launches Project Glasswing, says Claude Mythos is too dangerous to release publicly — TechCrunch

The trust problem at OpenAI

The New Yorker's eighteen-month investigation alleges Altman has a consistent pattern of deception around safety, citing memos from Sutskever and notes from Amodei. Meanwhile, OpenAI's own CFO is privately warning the Q4 IPO target is too aggressive. The company wants public-market credibility at the exact moment its credibility is under fire.

• New Yorker investigation paints Sam Altman as 'unconstrained by truth' as OpenAI nears IPO — Semafor
• OpenAI CFO privately warns 2026 IPO timeline is too aggressive, clashing with Altman — PYMNTS

The geopolitical split deepens

Z.AI's GLM-5.1 topped SWE-Bench Pro with 744B parameters, MIT-licensed, trained on Huawei chips with zero Nvidia hardware. That's exactly the capability the new Frontier Model Forum alliance is trying to contain. And Meta keeping its next-gen models Avocado and Mango proprietary before eventual open-source release signals that even the loudest open-source advocate has decided the frontier is worth guarding.

• Z.AI's GLM-5.1 becomes first Chinese model to top SWE-Bench Pro — trained entirely on Huawei chips — OfficeChai
• OpenAI, Anthropic, and Google form alliance to combat Chinese AI model copying — The Japan Times
• Meta plans hybrid open-closed approach for next-gen AI models Avocado and Mango — Axios

Agents get practical

Block's Managerbot proactively runs inventory, staffing, and marketing for Square sellers. Google quietly dropped an offline dictation app on iOS. Rocket is packaging AI as McKinsey-style consulting for a fraction of the price. The pattern: AI products doing specific jobs for specific people, shipping without fanfare.

• Block launches Managerbot, a proactive AI agent that runs Square sellers' businesses — VentureBeat
• Google quietly launches AI Edge Eloquent, a free offline dictation app for iOS — TechCrunch
• Rocket launches AI-powered McKinsey-style consulting at a fraction of the cost — TechCrunch

Infrastructure and the labour question

Eclipse Ventures closed $1.3B for physical AI and robotics. Aria Networks raised $125M for AI-native data centre switches. The hardware layer is attracting serious capital, and OpenAI's policy paper proposing robot taxes, a public wealth fund, and a 32-hour work week suggests even the builders know the downstream effects are coming. MIT Technology Review reports that economist consensus on AI job displacement has tipped, with 85,000 tech workers affected so far in 2026.

• Eclipse Ventures raises $1.3B for physical AI and robotics startups — TechCrunch
• Aria Networks raises $125M Series A for AI-native data center switches — SiliconANGLE
• OpenAI calls for robot taxes, public wealth fund, and 32-hour work week in policy blueprint — OpenAI
• AI job displacement shifts from Silicon Valley fear to economist consensus — MIT Technology Review

If you're building on frontier models, the message from today is clear: the window of open, anything-goes access is narrowing, and the companies providing your infrastructure are telling you they expect the rules to change.

---