What's this about?
OpenAI classified Astra as meeting the Critical cybersecurity threshold under its Preparedness Framework — the first model it has publicly labeled that way. The numbers: 100% on ExploitBench (up from 78.5% for Sol), 42.4% on ExploitGym, and 88.0% single-attempt on SRE-Bench binary reverse engineering versus Sol's 55.9%.
What exactly is the "Preparedness Framework"?
OpenAI's internal risk-tiering system: capabilities are tracked in categories (cyber, bio, persuasion…) and models crossing defined thresholds trigger mandatory safeguards before deployment. "Critical" is the top tier. Crossing it is why the shipping model refuses exploit-writing tasks and why the raw capability sits behind the vetted-access Daybreak program.
Why is this worth noting?
A frontier lab is now saying, in a launch post, that its flagship model has offensive cyber capability at a level requiring maximum internal controls. Every number is OpenAI's own evaluation of its own model — no independent verification — but the classification itself is a fact, and it explains the whole gated-access architecture the industry converged on this month.