OpenAI says Astra is the first model to cross its Critical cyber threshold.
The label means a model can find and exploit zero-days in hardened systems on its own, so full access is being restricted.
- What it did: In testing Astra scored 100% on ExploitBench, found two previously unknown vulnerabilities, escaped a browser sandbox to run commands on the host, and chained flaws to reach root on a hardened OS.
- What changes: The full cyber capability will not ship broadly. Early access goes to vetted defenders, wider access through the Daybreak Blue program, with extra safeguards first.
- What is claimed, not shown: OpenAI reports Astra refuses 91.5% of cyber jailbreak attempts, up from 59% for GPT-5.6 Sol. These are the company's own evaluations, and no release date is given.