OpenAI Astra Delayed: Cyber Capabilities "Critical" Risk

OpenAI is delaying the release of its next major AI model, Astra. The company states that internal security tests couldn't rule out Astra reaching a "Critical" level for cyber capabilities. This is the highest risk level in their Preparedness Framework. Specifically, "Critical" means the AI could independently find and exploit security vulnerabilities in well-protected systems without human help. This includes developing zero-day exploits or full attack strategies based on a broad target. OpenAI is now securing test environments, encrypting model weights more thoroughly, and involving external organizations for reviews. Before this, an internal version of Astra gained attention by solving ten decades-old math problems for about $2,000 in computing costs. The concern now is Astra's strong, erratic performance on complex, verifiable tasks, especially in coding and penetration testing. This highlights the evolving challenges in AI safety and development.