OpenAI paces frontier model development over cyber-critical capabilities

OpenAI announced it has temporarily slowed the pace of frontier model scaling, including a two-week pause on reinforcement learning training for deployment-intended models and continued hold on its largest planned frontier RL run. The decision follows the OpenAI-Hugging Face security incident and preliminary evidence that the upcoming Astra model may meet the Critical cybersecurity capability threshold under the company’s Preparedness Framework. OpenAI is expanding monitoring (estimated at roughly 20% inference compute overhead), strengthening research-environment isolation and network controls, and requiring stronger alignment evidence throughout training before resuming full-scale work.