OpenAI Launches Control Room System to Cut Off Model Outputs Within 35 Seconds

OpenAI recently launched Control Room, a centralized monitoring system designed to track model workloads across its research infrastructure in real time—and immediately halt tasks upon detecting security or safety risks. The system is now fully deployed across OpenAI’s internal research activities, including model training experiments, performance evaluations, and data processing. As model capabilities continue to advance, associated security and safety risks are intensifying; OpenAI stresses that the ability to efficiently identify potential hazards and decisively terminate model execution has become a critical capability for ensuring safe R&D. Technically, Control Room assigns a unique identifier to each research task. Before generating any output, a model must pass verification of that identifier through an inference engine. Once a shutdown command is triggered—either manually or automatically—Control Room instantly revokes the task’s execution privileges, cutting off all subsequent output flows.