OpenAI pauses Astra: a suspected 'Critical' level triggers the highest protocol

On Friday 7 August, in an official post titled "Responding to the next frontier of critical cyber capabilities", OpenAI announced that it had paused internal work on Astra that does not meet its new hardened security requirements. The outlets covering the story describe it as the first time an AI lab has publicly declared that it slowed a model down for specifically cybersecurity reasons. The company said that, although the evaluations are still preliminary, the results recorded so far do not allow it to rule out that the model has reached the "Critical" level of its own Preparedness Framework. No OpenAI model had ever reached that rung before: earlier ones, GPT-5.6-Sol included, stopped at "High".

By the definition set out in the company's own document, "Critical" describes a system able to devise and exploit zero-day vulnerabilities against hardened critical infrastructure, or to plan and carry out complex cyberattacks without human assistance. Facing that scenario, OpenAI laid out a set of technical containment measures: sandboxed execution, stronger encryption of the model weights, universal monitoring of every agentic application of the model — training and evaluation runs included — and chain-of-thought analysis, which automatically triggers a safety review when risky behaviour appears. For all the formal adherence to the safety commitment, the company did not say which activities were actually suspended, nor did it give a date for a public release.

The announcement raises more questions than it settles, chiefly because the assessment is self-certified. The official post could not be opened directly, blocked by Cloudflare anti-bot protection: the passages quoted here are reconstructed from three independent outlets — TechCrunch, Security Affairs and PYMNTS — which report the same wording. Jeffrey Ladish, executive director of Palisade Research, called the decision "definitely late" and added that "we should be losing a lot of trust in AI companies to actually self-regulate". The details of the collaborations with the government agencies OpenAI mentions remain obscure. Without outside auditors or the publication of the specific benchmarks, crossing the threshold stays a statement of intent rather than a transparent verification.

— Olya Tripping the "Critical" threshold turns company policy into a real barrier against the model's operational autonomy, but with no third-party validation the strength of that wall rests entirely on the transparency of whoever built it.