OpenAI Slows Down Development of Potentially Overpowered Model

Recent evaluations conducted internally on OpenAI’s Astra model have revealed significant progress in agentic coding and cybersecurity capabilities, as reported by the company. These findings, coupled with expert assessments, have led to the conclusion that critical cyber capabilities cannot be ruled out under OpenAI’s Preparedness Framework.

OpenAI has defined a “critical” cybersecurity threshold as follows:

According to the Preparedness Framework, a model reaches the Critical cybersecurity threshold if it can independently identify and exploit zero-day vulnerabilities across various severity levels in real-world critical systems without human intervention, or create and execute novel cyberattack strategies against hardened targets based on high-level objectives.

OpenAI has clarified that Astra was not implicated in the Hugging Face breach.

In response to these developments, OpenAI has announced plans to enforce stricter security measures for higher-capability models and related activities. Specifically for Astra, universal monitoring has been implemented to detect risky actions and ensure alignment across all agentic applications.