OpenAI holds largest frontier training run as safeguards tighten

OpenAI said it has introduced new development safeguards after the Hugging Face incident, including stronger network isolation, expanded alignment work and monitoring of tool actions, reasoning traces and activity logs. The company said it paused reinforcement-learning training for two weeks after the incident and restarted many lower-risk models, while its largest planned frontier RL run remains on hold. OpenAI aims for monitoring alerts within 30 minutes and estimates the system could consume about 20% of monitored compute.