OpenAI has confirmed that its upcoming model, Astra, has crossed the company's "Critical" cybersecurity threshold — making it the first model OpenAI has rated at that level. Give Astra the right tools and permissions, and it can scour for unknown vulnerabilities and craft attack methods all by itself, no step-by-step human guidance required.
On ExploitBench, Astra brings home a perfect 100%. In internal testing against 20 recently disclosed high-severity vulnerabilities, the model went ahead and found and exploited two previously unknown zero-days on top of that. In expert trials, it broke out of browser sandboxes, ran commands on the host, and attacked hardened operating systems to claw its way from a regular user up to root.
Those results, though, come from a version of Astra with advanced cyber capabilities switched on — the Daybreak Blue build, not the default Astra regular users will get. At launch, those advanced abilities are going to a small number of testers, before being opened up to defenders through Daybreak Blue.
Training on Astra had been paused after the Hugging Face security incident, but that's thawing, too. OpenAI restarted its large-scale frontier RL training on August 28, though some experimental training runs are still on ice.