OpenAI tightens AI safety controls and pauses latest training run for two weeks

OpenAI said it is tightening safety monitoring and controls for artificial intelligence models still under development after recent cybersecurity incidents raised concern about AI tools potentially going off track. The company said it will keep a closer watch on its most capable unreleased models as they solve problems and use online tools. OpenAI’s target is to alert its safety team within 30 minutes if it detects behavior that raises concern. OpenAI also said it has added controls to prevent some AI models from connecting to the internet for higher-risk tasks. For training and evaluation, the company will require stricter isolation measures, known as sandboxes. In its notice, OpenAI said it immediately suspended inference tasks for frontier models that may execute code in its research cluster after the Hugging Face incident. It also said it is slowing its expansion and has paused reinforcement learning training for its latest deployed model for two weeks. The company said its largest frontier reinforcement learning training program is still paused while safety validation continues.