OpenAI said it temporarily slowed the pace of scaling its models after a security incident involving Hugging Face and early signs that an upcoming model, Astra, may reach the critical cybersecurity capability level under its own risk framework. It paused reinforcement learning training on deployment-bound models for two weeks, put its largest planned frontier training run on hold, and tightened isolation and monitoring in its research environments. The company said it will revise its Preparedness Framework to cover both training and deployment.

• two-week pause in RL training
• monitoring overhead ~20% of inference compute
• 30-minute alert target
• Astra assessed 7 August 2026