Pacing Model Development in an Era of Cyber-Critical Capabilities

OpenAI has temporarily slowed frontier model development after two triggers: a security incident involving Hugging Face and preliminary evidence that its upcoming model, Astra, may meet the threshold for critical cybersecurity capabilities under its Preparedness Framework. The company paused two weeks of reinforcement learning training and has kept its largest planned frontier RL run on hold while conducting smaller-scale evaluations. OpenAI has implemented three reinforcing safeguards across research and deployment: enhanced monitoring with 30-minute alert windows, stronger alignment requirements throughout training, and tighter security controls including workload isolation and network segmentation. Monitoring overhead runs approximately 20% of inference compute. The company acknowledges these measures represent a broader approach that extends beyond its existing Preparedness Framework, and views stronger evidence of aligned behavior as a prerequisite before scaling resumes.