OpenAI Pauses Some Internal Work as Astra Approaches a 'Critical' Cyber Bar
Preliminary assessment can't rule out Astra reaching 'critical' cyber capability, so OpenAI has paused internal activities that lack enhanced safeguards.
OpenAI has paused part of its internal work because its unreleased Astra model may be too capable at cybersecurity for its own safeguards to handle.
What happened
In a statement on August 7, OpenAI said internal and expert evaluations show Astra making significant progress in agentic coding and cybersecurity capabilities. Under the company's Preparedness Framework — its staged system for assessing dangerous capabilities — the assessment concluded that OpenAI "cannot rule out" Astra reaching the critical cyber capability threshold.
The framework defines that threshold concretely: a model that can, without human intervention, identify and develop working zero-day exploits across many hardened real-world critical systems — or devise and execute end-to-end novel cyber attack strategies from only a broad objective. According to coverage, this is the first time a model has neared the critical level.
OpenAI says it has expanded robustness testing, tightened security controls on Astra-related work, and paused internal activities that do not yet meet the enhanced requirements. It also clarified that Astra was not involved in the reported attack on Hugging Face.
Why this matters beyond OpenAI
Capability thresholds used to be theoretical documents. Here they are operating: a company stopped parts of its own project because a system got close to a line drawn in a policy document. That is the AI safety regime becoming operational — and it will happen more often as agents get better at writing and fixing code.
The question the industry now has to answer is the uncomfortable one: what happens when a company's most capable system is also the one its competitors most want, and the safeguard pause is expensive? A framework is only as strong as the willingness to keep the pause.
The moment a lab pauses its own most impressive project, safety stops being a press release and becomes a cost.