Is OpenAI's next model too dangerous to ship? Maybe.
OpenAI's latest safety testing has reportedly found that its upcoming AI model, Astra, carries cyber capabilities that have hit the company's highest risk tier — one level above GPT-5.6 Sol.
At that level, OpenAI's own standards suggest the model could operate without a human in the loop: finding and exploiting zero-day vulnerabilities in critical systems, and carrying out the entire chain from target selection to attack design and execution on its own.
That's apparently enough to make OpenAI pump the brakes. The company has suspended parts of Astra's internal testing and tightened restrictions on internet access, tool use, and model weight permissions. Next up, it plans to hand the model to government agencies and outside safety organizations for further review.
Earlier reports had pointed to Astra launching as soon as next week. Now, that timeline is looking a lot less certain.