OpenAI's internal evaluation of its forthcoming AI model Astra indicates that it may autonomously develop zero-day exploits against hardened real-world systems and independently design and execute cyberattacks from high-level goals, qualifying it for a 'critical' cybersecurity risk tier under OpenAI's Preparedness Framework. This represents a significant advancement beyond the GPT-5.6-Sol model, which was rated 'high' risk. To mitigate risks, OpenAI has implemented strict development controls including isolated testing setups, network restrictions, model weight protections, and universal monitoring to intercept high-risk behaviors. Development projects not adhering to these controls are paused. Astra is unreleased and was not involved in recent AI-driven hacking incidents. OpenAI plans to test Astra with government agencies and AI safety groups and share security recommendations with third parties.

The model Astra's capabilities to autonomously create zero-day exploits and conduct end-to-end cyberattacks pose a critical cybersecurity risk if misused. This elevates the threat landscape by potentially enabling AI-driven autonomous cyberattacks without human intervention. However, Astra is currently unreleased and under strict internal controls to prevent misuse. No known exploits in the wild involve Astra at this time.

OpenAI has paused any internal development involving Astra that does not comply with newly mandated security controls, including isolated testing environments, strict network restrictions, and enhanced model weight protections. Universal monitoring is deployed to detect and shut down high-risk or misaligned behavior automatically. External testing will be conducted alongside government and AI safety organizations to further evaluate and manage risks. Since Astra is unreleased and tightly controlled, no immediate external mitigation actions are required.