OpenAI unveils astra, a cybersecurity model with advanced capabilities and enhanced safeguards

OpenAI has announced the development of Astra, a model that meets the Critical cybersecurity capability threshold under its Preparedness Framework. Astra can identify and exploit previously unknown vulnerabilities across well-protected systems without human guidance. The model was tested on ExploitBench, achieving a perfect score, and on an internal benchmark with 20 high-severity V8 vulnerabilities, where it outperformed GPT-5.6 Sol in code-execution rates. Astra also discovered and used two zero-day vulnerabilities in an exploit chain. To mitigate risks, OpenAI has implemented stronger safeguards, including refusal mechanisms for harmful requests, additional protections against misuse, and monitoring to stop unauthorized activity. These measures were reinforced following the OpenAI-Hugging Face incident. Astra’s capabilities will initially be available to a limited group of testers through Daybreak Blue, with expanded access for defensive use. OpenAI plans to share detailed safety and security evaluations in Astra’s system card at launch.