OpenAI to Give Partners Early Access to Astra, Its First 'Critical' Cyber Model
In one line: OpenAI reportedly plans to give select partners early access to Astra, the first model it classifies as "Critical" for cyber capabilities.
Key points
- According to Wired, OpenAI has, for the first time, rated a model — Astra — as "Critical" for cyber capability under its internal risk framework.
- The company will reportedly grant early access to select partners before wider release, giving them time to shore up their defenses.
- A "Critical" rating signals OpenAI's own assessment that the model could be meaningfully misused for cyberattacks or vulnerability discovery.
Why it matters
It appears to be the first time a developer has labeled its own model's offensive cyber abilities as "Critical," and the staggered, defender-first rollout could set a precedent for frontier-model safety practices. At the same time, the same capabilities carry real misuse risk, reviving the tension between shipping powerful models and containing security fallout.