OpenAI rates Astra "Critical" for cyberattacks and plans to release it

OpenAI says its next model can find and exploit unknown security holes on its own, and it is releasing it anyway with the locks on.

The designation

On September 1, 2026 OpenAI published “Path to Astra.” It now believes Astra meets the Critical cybersecurity threshold of its Preparedness Framework, the first model it has ever designated at that level. Critical means the model “can find previously unknown security flaws and develop ways to exploit them across many well-protected systems without a person guiding each step.” In tests it built a full browser-compromise chain that escaped the sandbox, escalated from an unprivileged user to root on a hardened operating system, and used two zero-days it discovered along the way, which OpenAI is now disclosing to the maintainers.

The release plan

OpenAI says Astra’s safeguards “sufficiently minimize the risk of severe harm for release,” so the model is coming “soon,” but the advanced cyber work goes first to a small group of alpha testers and then through the Daybreak Blue defensive program. The company also admits the safety checks will “occasionally flag legitimate activity” and pause or stop tasks, including ones unrelated to security.

The silent reasoning

There is a second story underneath. The Information reported, per Techmeme, that Astra uses “recurrent depth,” a technique that reuses transformer layers so more of the reasoning happens in latent activations instead of readable text. It makes the model cheaper and stronger, and harder to monitor. OpenAI’s chief scientist Jakub Pachocki reportedly answered that the effective depth stays within about twice GPT-4’s and that chain-of-thought monitoring is preserved.

Astra’s earlier milestones, ten solved math problems in early August and a pause of its training a week later, belong to the same model. It now arrives with a hacker’s skill set and a partly silent mind, and the release will test whether safeguards can keep pace with capability at this level.