OpenAI built a 35-second kill switch for frontier AI
OpenAI just introduced Control Room, a centralized system built to monitor research workloads and pull the emergency brake when safety concerns arise. Hooked into research inference and agentic sampling tasks, it can shut down model output within 35 seconds of an intervention decision. It is designed to keep humans firmly in control as frontier models grow more capable.