Responding to Next-Generation Core Cyber Capabilities

OpenAI stated in a recent internal evaluation of its next model Astra that agentic coding and cybersecurity capabilities have significantly improved. Based on expert assessments, it determined that it cannot rule out the possibility that Astra has reached Critical cyber capabilities as defined in its Preparedness Framework.

In the Preparedness Framework, the Critical threshold is the ability to identify and develop functional zero-day vulnerabilities of all severity levels across various security-hardened real-world critical systems without human intervention, or to design and execute new cyberattack strategies from start to finish against security-hardened targets when given only high-level attack objectives.

OpenAI is continuing its evaluations while strengthening safeguards to match this level of capability.

• It applies isolated test environments, restricted network and tool access, encrypted model weights, enhanced monitoring and detection, and sandbox execution to high-performance models and related activities.
• It suspends internal Astra-related activities that do not meet enhanced security control requirements.
• It applies general monitoring to all Astra agentic applications and training/evaluation processes to detect risky behaviors and alignment failures.
• The monitors evaluate the model's Chain of Thought and trigger security reviews and halt procedures if high-risk activities are detected.
• It collaborates with relevant government agencies, some AI safety organizations, and third-party testing partners to safely conduct high-risk evaluations.

OpenAI added that Astra is not the model involved in the Hugging Face breach. It also stated that models with advanced cyber capabilities should be used for defensive purposes, finding vulnerabilities before attacks, and announced it would work with governments, safety research institutions, and civil society to promote responsible deployment.