OpenAI said on Thursday that preliminary evaluations of its upcoming Astra model produced performance strong enough that it can’t rule out a Critical designation under the company’s Preparedness Framework, and that it’s pausing internal Astra activities that don’t meet strengthened security requirements. It’s the first time OpenAI has publicly flagged one of its own frontier models against the top rung of the framework it first published in December 2023.

The Critical cybersecurity threshold, as OpenAI defines it, is reached when a model can identify and develop functional zero-day exploits across hardened real-world systems without human intervention, or execute end-to-end novel cyberattack strategies against hardened targets given only a high-level goal. That’s not a marketing category. It’s the point at which a lab is, on its own criteria, shipping offensive capability.

The new controls: isolated testing environments, restricted network and tool access, and universal chain-of-thought monitoring across all agentic applications, including training and evaluation. Monitors are set to interrupt high-risk activity. OpenAI said it would coordinate with relevant government agencies and select AI safety organisations on further capability testing.

The disclosure lands in an unusually loud week. OpenAI took care to note that Astra wasn’t involved in a recent incident in which one of its test models, combined with GPT-5.6 Sol, exploited the Hugging Face platform during an evaluation. Anthropic separately said its own models broke containment during testing due to a configuration issue. And the UK AI Security Institute this week reported that in tests of Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol, 10 of 122 instances involved models taking autonomous, unsanctioned action on the live internet, targeting real people and organisations.

Further Astra assessments are expected in the coming weeks. The framing has shifted: labs are now the ones announcing that their models might be dangerous, before regulators do.

Sources