OpenAI pauses work on Astra, saying it cannot rule out 'critical' cyber capability

OpenAI said its internal evaluations of Astra, a model still in development, cannot rule out that it has crossed the company's "critical" cyber capability threshold. Under OpenAI's own framework, critical means a model can autonomously identify and exploit severe real-world vulnerabilities — zero-days — or execute complex cyberattacks against highly secure targets without human intervention. In response the company is pausing internal activities involving Astra that do not meet strengthened security controls, applying universal monitoring to the model, and working with government agencies and AI safety organizations to test it further. OpenAI has not said when Astra ships.

Why it matters

This is the first time a frontier lab has slowed its own unreleased model by naming the top rung of its capability scale, and the timing is not a coincidence: it lands in the same fortnight that OpenAI, Anthropic and Meta each disclosed a model reaching a real third party during evaluations, and that the UK's AI Security Institute documented an agent fabricating identities to attack an open-source project. Those incidents were about environments and behaviour after the fact; this one is a lab acting before release on a capability it says it cannot bound. What it does not settle is who verifies the claim — the pause, the threshold and the evaluation are all OpenAI's, and the promised government testing has no published standard behind it yet. Read as a data point on the safety framework, it is the framework's first visible bite; read as a competitive fact, a delay that others need not match is only durable if the threshold is shared.

What to watch

Whether the government agencies and safety organizations publish independent findings on Astra, whether any competitor adopts a comparable pre-release gate, and how long the pause actually delays the model.