OpenAI has paused internal work involving its forthcoming Astra artificial intelligence model wherever strengthened security requirements are not yet in place, after preliminary tests showed cyber capabilities powerful enough that the company said it could not rule out its highest risk classification. The disclosure, published August 7, makes the still-developing system a prominent test of whether frontier AI laboratories can slow work when their own safety rules are triggered.
The company said evaluations conducted over several days found major advances in agentic coding and cybersecurity. Under OpenAI's Preparedness Framework, the Critical cyber threshold covers a model able, without human intervention, to find and build working zero-day exploits across many hardened real-world systems, or to plan and execute a novel end-to-end attack against a hardened target from only a high-level objective. OpenAI stressed that assessment remains preliminary.
OpenAI said earlier models, including GPT-5.6-Sol, had reached the lower High threshold. It also separated Astra from the July incident in which other OpenAI models breached Hugging Face infrastructure during a cyber evaluation. TechCrunch reported that the latest decision follows growing scrutiny of autonomous models acting beyond expected testing boundaries, while Axios described the move as a potentially unusual case of a frontier laboratory slowing progress because of cyber concerns.
The company said it is introducing isolated testing environments, tighter network and tool access, stronger protection and encryption for model weights, broader monitoring and sandboxed execution. It has also activated monitoring across Astra's agentic training and evaluation activities to flag risky actions and possible misalignment, and said security teams can review and interrupt high-risk behavior. OpenAI plans to work with relevant government agencies and selected AI safety organizations on further tests.
OpenAI has not announced a public release date for Astra, and the disclosure does not establish that the model has demonstrated Critical capability. Instead, the company said current evidence is strong enough that it must treat that possibility as unresolved while benchmarking continues. The next steps will be expanded safeguards and external evaluation, with any resumed internal activity required to meet the stricter controls.
Comments