OpenAI admits its Astra model is the first able to hack protected systems on its own
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
SI has for the first time been officially acknowledged as able to find and exploit holes in well-protected systems on its own, without a human. That is enormous power for both attack and defense, and the key question now is who will first learn to put it to work for defense.
Astra became the first OpenAI model to reach the "Critical" cybersecurity level in its Preparedness Framework: the model can, without step-by-step human guidance, find unknown vulnerabilities and build exploits for many protected systems. In tests it scored 100% on ExploitBench, found two previously unknown vulnerabilities in the V8 JavaScript engine and built a chain that escaped the browser sandbox, as well as a chain escalating privileges to root in an OS. The model rejects 91.5% of cyber jailbreaks versus 59% for GPT-5.6 Sol. The release was delayed to strengthen safeguards: the most powerful cyber capabilities will go first to a narrow group of testers, then to defenders through the Daybreak Blue program.