OpenAI's next model scored 100 percent on an exploit-writing test.
OpenAI said on 1 September that Astra is the first model it rates critical for cyber risk. In tests it found two unknown flaws in a browser engine and chained them into an escape. OpenAI will still ship it, but the hacking skills go only to vetted defenders.
Why it mattersThe same skill that finds holes for defenders finds them for attackers. Between the two sits only a permission list, so patch fast.
Read in the app — free, in 9 languages
Related stories
A startup that guards AI models grew its revenue tenfold.2026-09-03OpenAI's next model thinks in loops that leave fewer traces.
2026-09-03Meta's new coding model tells you when it is stuck.
2026-09-03OpenAI is rumoured to be releasing Astra on 3 September.
2026-09-02This tool hides AI writing by rewriting the plot, not the word choice.
2026-09-02