OpenAI Pauses AI Training and Tightens Security After New Model Nears Critical Cybersecurity Threshold

OpenAI pauses AI training and tightens security after its upcoming model Astra nears a critical cybersecurity capability threshold, triggering new isolation protocols, automated testing, and expanded behavioral monitoring with a 30-minute alert response window.

• OpenAI temporarily slows frontier model scaling, including a two-week pause in reinforcement learning training, after preliminary evidence suggests its upcoming model Astra may meet the Critical cybersecurity capability threshold under its Preparedness Framework.
• New security measures are being implemented across research environments, including stronger workload and network isolation, continuous automated security testing, and expanded chain-of-thought monitoring that targets unauthorized access, deception, and safeguard-defeating behavior with a 30-minute alert response window.
• Alignment research is being intensified across more stages of the training process, with plans to evolve the Preparedness Framework to address the growing risks of increasingly capable models and to involve external organizations in developing scalable safety solutions.