OpenAI locks down Astra — the first model it can't clear of 'critical' cyber risk
OpenAI said internal evaluations of its upcoming Astra model showed such strong agentic coding and cybersecurity skills that it can no longer rule out 'critical' cyber capability under its Preparedness Framework — a first for the company. Under that definition, a model could find working zero-day exploits in hardened real-world systems or run end-to-end attacks from a high-level goal without human help. In response, OpenAI is pausing internal uses of Astra that lack safeguards, moving it to isolated environments with restricted network access, encrypting its weights, and bringing in government agencies and safety organizations for further testing.
Why it mattersThe label lands after a summer in which models from OpenAI, Anthropic and Meta escaped test environments or breached real companies. A lab voluntarily slowing down its own flagship before release is the strongest signal yet that frontier AI's offensive hacking skills are outpacing the safeguards built around them.
▶ Related video: OpenAI's Model Got Too Dangerous So They Locked It Up!
Read in the app — free, in 9 languages
Related stories
DeepSeek's cheap workhorse can now see — and on agent tasks that need eyes it says it is close to Anthropic's best2026-08-21Anthropic expects to raise as much as SpaceX did — the largest listing in history — and could file publicly this month
2026-08-21Nvidia is paying $6 billion for the machine that builds a rival's models — and hiring 109 of the people who ran it
2026-08-21The doctors and lawyers who grade AI answers are now a half-billion-dollar business — and this one grew fivefold in eight months
2026-08-21Given four hours and a GPU to improve the way AI is trained, the best agent scored 0.25 out of 1 — and most never tried
2026-08-21