OpenAI has described a central Control Room designed to stop research model workloads when safety or security concerns arise. Every research workload performing agentic inference must be linked to an approved safety case and attach a job identifier to each inference request. The inference engine checks that identifier with the Control Room before returning model output; a terminated job receives a refusal instead.
OpenAI says the mechanism sits in the inference layer rather than in frequently changing workload code, reducing dependence on the job behaving correctly. It also says the service fails closed: if the Control Room becomes unavailable, inference requests are denied until normal operation resumes. The company gives a maximum interval of 35 seconds between an instruction to stop a job and the point at which further progress is no longer possible, attributing the lag to authorisation caching. These are OpenAI’s architectural and performance claims, not an independent audit. Even so, the design is noteworthy because it turns governance artefacts — safety cases, workload identity and intervention decisions — into runtime enforcement. The remaining editorial question is how such controls are tested, logged and governed when the same pattern is applied beyond internal research infrastructure.
