Path to Astra: critical capabilities and frontier safeguards
Astra is OpenAI's first model rated to find unknown security vulnerabilities and develop working exploit chains across hardened systems without human guidance.
- Astra achieved a 100% score on ExploitBench and discovered two zero-day vulnerabilities during testing on recent V8 engine flaws.
- The model refused 91.5% of disallowed cyber requests in jailbreak evaluations, compared to 59% for GPT-5.6 Sol.
- OpenAI restarted a previously paused large frontier reinforcement learning run for Astra on August 28 following training-environment security upgrades.
- Access to advanced cybersecurity capabilities will initially launch to a small alpha group before widening through Daybreak Blue.
Defenders and developers using Astra will face stricter access tiers and active monitoring that can pause or stop tasks flagged as potential cyber misuse.
