OpenAI is about to ship its first model that it rates critical for cyber capabilities, and it's giving a few partners a head start to prepare.
The company plans to release a new model called Astra. It's the first model OpenAI has assigned its critical risk tier for cyber skills. Ahead of the wider launch, OpenAI is giving select partners early access to Astra so they can shore up their own defenses. No previous OpenAI model has carried this rating for cyber capabilities.
A critical cyber rating signals the model's abilities could meaningfully help with serious intrusion or exploitation work, not just describe it in the abstract. That's a shift in posture: offensive cyber skill is becoming a real safety line item for frontier labs, not a hypothetical for some future release. Handing partners a warning window before broad release amounts to OpenAI admitting capability is outrunning most organizations' readiness to defend against it.
Worth remembering: 'critical' here is a label OpenAI assigned to its own model. The company is simultaneously the rater and the vendor, which is worth keeping in mind as the industry leans on self-reported safety tiers instead of independent ones.