OpenAI Launches Central Control System 'Control Room' That Can Halt Model Output Within 35 Seconds

Nov 10 — OpenAI posted on X that it has built a central system called Control Room, used to track model workloads across its research infrastructure and to quickly halt them when safety hazards arise. The system covers all research inference workloads, including agent sampling for model training and evaluation as well as data processing, and operates in tandem with its misalignment monitoring system. It can stop model output within 35 seconds of an intervention decision being made, in order to terminate or pause the relevant tasks. OpenAI said Control Room is part of its overall approach to safely training and deploying frontier models.