Update
OpenAI introduces 35‑second emergency brake for research AI output
OpenAI announced its Control Room system can stop new model output within 35 seconds of a stop instruction. The cutoff is enforced by the inference engine, independent of the research code, which may continue running. The process begins when Control Room receives the instruction; detection and decision time are excluded from the 35‑second limit. The system is already deployed across OpenAI’s research infrastructure and uses job IDs to manage authorizations. OpenAI said the design centralizes enforcement and that it continues to search for workloads missing Control Room checks.