OpenAI says Astra model may reach critical cyber capability threshold

OpenAI has said its upcoming artificial intelligence model, Astra, has shown significant advances in agentic coding and cybersecurity. The company stated it cannot rule out the model reaching the 'Critical' cybersecurity capability threshold under its internal Preparedness Framework.

OpenAI's internal evaluations, conducted over recent days with expert assessments, showed a notable improvement in Astra's ability to perform cybersecurity tasks. Under the framework, a model reaches the Critical threshold if it can identify and develop functional zero-day exploits across multiple hardened, real-world critical systems without human intervention, and develop novel cyberattack strategies based only on a high-level objective.

OpenAI said Astra was not involved in the exploitation of Hugging Face. The company has expanded robustness testing of Astra's safeguards, implemented stricter security measures including isolated testing environments and encrypted model weights, and paused internal Astra activities that do not meet strengthened security requirements. OpenAI plans to work with government agencies and selected AI safety organisations to test Astra's capabilities.