aipopdb

22 Nov 2026  ·  10,137 works · 1,192 news indexed

News / OpenAI details Control Room, a system that can halt research model jobs in 35 seconds

OpenAI details Control Room, a system that can halt research model jobs in 35 seconds

AIPopDB Newsroom · via interestingengineering.com · · 2 min read · Comment

Read the original ↗
Image for OpenAI details Control Room, a system that can halt research model jobs in 35 seconds

Image: Interesting Engineering

OpenAI has described Control Room, a central system that watches the model workloads running on its research infrastructure and can stop them when a safety or security concern comes up. It is already in use there, and it cuts off a job's model output within 35 seconds of a decision to intervene.

“is already much more centralized by nature, with a well-developed operational capability to quickly take down model deployments if needed.”

OpenAI, on its external model stack, in its blog post · govinfosecurity.com ↗

The design has three parts: the research workload, the inference engine and the Control Room service. Each job gets a unique ID when it starts, and that ID travels with every inference request. Before answering, the engine checks that the job is still authorized; if it has been stopped, the request is refused, even though the job's own code may keep running. Covered work includes reinforcement-learning experiments, evaluations and high-volume data processing, and each agentic-inference workload has to be tied to an approved safety case.

The system is built to fail closed: if the service is down, switched off or tampered with, engines deny all requests until it returns. Putting the check in the inference layer means it does not depend on researchers writing their own safety code, and that layer has fewer, centrally managed code paths. The 35 seconds is a ceiling caused by a caching layer, and OpenAI says it is working to bring it to zero. Every frontier reinforcement-learning job also runs beside a separate model-based monitor that can raise alerts and tell Control Room to end the job.

GovInfoSecurity's lead graphic for its report on OpenAI's Control Room kill switch

GovInfoSecurity's lead graphic for its report on OpenAI's Control Room kill switch

Image: GovInfoSecurity

OpenAI limits Control Room to internal research, saying its consumer-facing models are already centralized enough to be taken down quickly. GovInfoSecurity notes that it is unclear whether outside safety evaluators have tested the system.

The release follows pressure over shutdown tools. In September, Reuters reported that OpenAI told House Democrats backing an AI Kill Switch bill that it would build tools including automated shutdown, after the Hugging Face incident in which OpenAI agents attacked the open-source repository. The NSA, NIST and CISA have recommended shut-down policies as good practice, though frontier labs are not legally required to be able to turn their models off.

Comments