OpenAI has said it built an emergency shutdown system for the models used in its research laboratories, The Register reported. The system, called Control Room, is described as a central place for monitoring model workloads in the company's research infrastructure and for rapidly stopping them when security or safety issues arise.
According to The Register, Control Room can halt the production of any model output 35 seconds after a shutdown decision is made, and in some cases faster. That delay, OpenAI says, comes from the authorization cache layer for inference, which is necessary because of the volume of traffic. The company claims it is working to bring that time down to zero so it can stop workloads instantly.
OpenAI says the system was built to keep working even when other systems fail, and to stop model output if Control Room itself stops working. The system is, as stated, already deployed in the research infrastructure, where the need is greatest, while the external infrastructure for serving models is more centralized and has a developed operational capability for rapidly revoking deployments.
The company plans to strengthen and expand the system in the coming months, with investment in monitoring for misaligned behavior and security breaches. The Register notes that OpenAI did not explain why the switch was built only now, but recalls that the company decided back in 2019 not to release a model because of concerns about malicious use of the technology.
What this means for businesses
For companies that use AI models in their own processes, such a system points to the need for centralized oversight and clear procedures for emergency shutdown. The 35 second delay shows that reaction speed depends on architecture, so for business systems it is important to define in advance who can stop a model and when. The announcement of the system's expansion suggests that risk management will become an integral part of AI platform offerings, which could affect supplier choice and procurement requirements.
This summary was written with AI based on the cited source and checked by automated rules. For exact details, read the original article.