OpenAI tightens security with new monitoring system

OpenAI announced new safety policies to control risks during model testing. The measures include more detailed monitoring during development and stronger alignment and safety efforts after training. OpenAI stressed that as model capabilities grow, risks in internal development rise, so its standards must stay ahead of those risks.

Earlier, OpenAI paused some reinforcement learning training for two weeks after a security incident on Hugging Face. Now, training for lower-risk models has restarted, but large frontier models remain paused. A new monitoring system will check tool actions, reasoning traces, and activity logs, aiming to alert on suspicious activity within 30 minutes. It uses about 20% of the computing power of the monitored process. OpenAI said it will share more technical details later, but has not yet published an official post-incident report.