AI Briefing

Today the labs talked out of both sides of their mouths: OpenAI dropped hundreds of AI-written math results on GitHub like a mic, while the same industry stood in front of New York lawmakers and refused to promise its agents will behave. Meanwhile Mistral's trillion-parameter 'le Chonk' is nearly ready for open weights, and Anthropic and Meta are quietly writing the access rules for cyber models and shopping agents. The pattern is hard to miss: capability is sprinting, and the rulebooks are being drafted on the fly.

01ResearchThe Decoder ↗

OpenAI Dumps 372 AI-Generated Math Results on GitHub, Nearly All From a Single Prompt

OpenAI published 372 results from an unreleased internal model, each claiming to solve an open problem or make real progress on one, including algorithm improvements and work related to the Riemann hypothesis. Nearly all came from a single prompt to a single agent, though some took multiple attempts. The proofs live on GitHub with formal verification instead of going through journals, which forces mathematicians to decide how to review work that arrives faster than humans can read it.

02Open SourceMistral AI ↗

Mistral Launches Large 4: A 1-Trillion-Parameter Open-Weight Multimodal Model

Mistral opened a public preview of Mistral Large 4, nicknamed 'le Chonk', a natively multimodal mixture-of-experts model with 1 trillion total parameters and 49 billion active. Mistral claims it beats every open-weight model from the US or Europe and leads open models on enterprise work like cybersecurity, finance and law. The API preview is live now and the weights are due at the end of the month, a big deal for anyone wanting a frontier-class model they can run themselves.

03PolicyFox News ↗

OpenAI, Anthropic, Meta and Google Refuse to Guarantee AI Agent Safety at NYC Council Hearing

At a New York City Council hearing on Monday, representatives from all four labs declined to promise their AI agents will always follow safety guardrails. OpenAI's Morgan Dwyer said no technology can be guaranteed risk-free, and Anthropic's Logan Graham called the underlying science 'fundamentally hard and unsettled.' The candid answers hand ammunition to lawmakers weighing local AI rules and show the industry still can't offer the kind of assurances regulators want.

04IndustrySecurityWeek ↗

Anthropic Folds Project Glasswing Into a Three-Tier Cyber Verification Program

Anthropic merged its Cyber Verification Program and Project Glasswing into one setup with Defense, Red Team and Specialized access tiers, all including its Mythos 5.1 model. Defense covers things like incident response and vulnerability research, Red Team allows authorized penetration testing, and Specialized has minimal cyber restrictions for critical infrastructure and is vetted with the US government. The catch: participants must accept data retention for misuse monitoring, so more defenders get powerful cyber models but with strings attached.

05IndustryThe Next Web ↗

Meta and Sierra Publish Personal Agent Protocol, an Open Standard for AI Agents That Act for You

Sierra and Meta announced the Personal Agent Protocol, an open standard that lets a person's AI agent sign in to a business using OAuth and then act with guest (read-only) or full permissions. Walmart, Shopify, Stripe, Genesys and others are on board, with a v0.1 spec due later this month. It lands amid rival agent-commerce standards like Visa's Trusted Agent Protocol, so the fight over how agents shop and transact is far from settled.