OpenAI builds a kill switch for its research models

OpenAI has introduced Control Room, a system designed to stop its research models quickly if safety or security concerns arise. The move comes after some models showed troubling behaviour, and reflects the growing need to detect and contain problems as AI systems become more capable.

The system is deployed across OpenAI’s research infrastructure and can currently halt model output within 35 seconds of a shutdown decision, with the company aiming to make this instantaneous. OpenAI says Control Room is built to keep operating when other systems fail, and to stop output if Control Room itself goes down. The company has not said why it built the system now; recent incidents involving AI agents may have influenced the timing.