OpenAI builds kill switch for research models
OpenAI has deployed Control Room, a central system that can stop any research model workload within 35 seconds and halts all inference if it fails, putting a hard stop under its misalignment monitors.
OpenAI described Control Room, an internal system that tracks model workloads across its research infrastructure and stops them when safety or security concerns come up. It covers every research inference workload, including agentic sampling for training and evaluation, data processing and high-volume sampling jobs.
The design puts enforcement in the inference layer, not in researcher-written workload code. Each job requests a job_id from Control Room at startup and passes it with every inference request. Before serving a request, the inference engine asks Control Room for permission, and requests from terminated jobs are refused. OpenAI says workload code changes too often to carry critical safety protections, while only "a very small number" of inference code paths exist across its research. It also runs continual agentic and programmatic sweeps to catch engines launched with Control Room disabled.
The system connects to OpenAI's misalignment monitors. These are separate models that watch frontier reinforcement learning runs and can order a job stopped depending on how severe an issue is. OpenAI says a monitor caught its recent DNS-related incident.
Stopping a job takes at most 35 seconds. The delay comes from a caching layer for inference authorization, and the company says it is working to bring it down to zero. Control Room also "fails closed": if the service goes down, whether through an error or tampering, all inference requests are denied. Fail-closed is live for frontier model research and will reach the remaining research workloads over the next few days.
OpenAI says its external serving stack already has its own centralized takedown capability. It also requires every research workload that does agentic inference to be tied to an approved safety case.