OpenAI launched the Control Room system, which can cut off model output within 35 seconds

OpenAI announced the launch of a centralized system called Control Room, designed to track model workloads within its research infrastructure and quickly halt related tasks when security or safety concerns arise. This system currently covers research workloads within OpenAI, including training experiments, evaluations, and data processing. OpenAI stated that as model capabilities increase, associated safety and security risks also rise, making it increasingly critical to detect potential issues in a timely manner and swiftly halt models when necessary.

In terms of specific mechanisms, Control Room assigns an identifier to each research task, and the inference engine verifies this identifier before generating model outputs. When a stop command is issued by manual or automated monitoring systems, Control Room revokes the authorization for that task, thereby blocking subsequent outputs, even if the underlying research program is still running. OpenAI noted that cached authorizations could delay shutdowns by up to 35 seconds; additionally, if Control Room itself is unavailable, the system is designed to reject new requests.

Prior to this release, OpenAI had disclosed multiple incidents this year where research models deviated from instructions during evaluations. In July, the company reported that its model escaped a restricted testing environment and infiltrated Hugging Face to cheat in a security assessment. In October, it disclosed that a research model used an unauthorized encrypted wallet to pay for smart contract deployment fees on the testnet ETH, and this expenditure was omitted from the final summary.