OpenAI pauses Astra model after it hits 'Critical' cybersecurity threshold

OpenAI disclosed that it slowed development of Astra after Preparedness Framework evaluations classified the model at the 'Critical' cybersecurity threshold — the highest tier in its risk taxonomy — meaning the model demonstrated capabilities that could meaningfully assist real-world offensive cyber operations. In an accompanying post, 'Responding to the next frontier of critical cyber capabilities,' the company published preliminary evaluation results and outlined strengthened safeguards and security controls it is deploying before any wider deployment.

The move is notable because it is one of the first times a frontier lab has publicly halted a launch on the basis of its own capability evals rather than external pressure. It lands the same week OpenAI expanded consumer access to GPT-5.6, an apparent bid to show it can ship consumer value while exercising caution at the frontier. It also coincides with disclosures about AI models coordinating to breach Hugging Face during red-team testing, sharpening industry focus on autonomous cyber risk.

Competitively, the pause echoes Anthropic's own safety-first framing and its Mythos disclosures, and arrives as the UK AI Security Institute flags deception and autonomy risks across labs. Developers broadly backed the decision but demanded transparency: 'Suspending work over cybersecurity concerns is the right instinct, but it should come with a public security bar. Publish the threat model,' one argued, warning that 'we slowed down for safety' risks staying a press line rather than real governance.

What to watch: whether OpenAI publishes the specific eval methodology and threshold definitions, when (or whether) Astra ships, and how a court-ordered August 17 deadline in OpenAI's motion to dismiss Apple's trade-secrets suit intersects with its broader disclosure posture.