OpenAI Reports Development Status of Next-Gen Model Astra, Enhances Safety Measures After Detecting High Cybersecurity Capabilities

It is stated that Astra may meet critical criteria within OpenAI's "Preparedness Framework" regarding cybersecurity capabilities.

Specifically, it was shown that Astra's ability to identify vulnerabilities and develop exploits has improved significantly compared to the previous model, GPT-5.6 Sol. It achieved a score of 100% on the "ExploitBench" benchmark and successfully discovered two zero-day vulnerabilities and utilized them in exploit chains.

Considering the risks associated with Astra's advanced capabilities, OpenAI has strengthened protection measures against cyberattacks and unauthorized operations. The company has postponed parts of the development and release process and introduced mechanisms to automatically review AI processing content.

During testing, no behavior was observed where Astra bypassed the restrictions of the automatic review. Furthermore, tests to verify the presence of misconduct showed that the probability of Astra engaging in such behavior is extremely low compared to GPT-5.6 Sol.

Astra is scheduled to be available soon, though it will initially be limited to a small group of testers. OpenAI stated that access will subsequently be made available through Daybreak Blue.