OpenAI launches Control Room system to monitor model workloads in research infrastructure

OpenAI has announced the deployment of a new centralized monitoring system called 'Control Room' to manage security and safety risks that may arise during the development of artificial intelligence models.

The Control Room system is designed to cover all research infrastructure workloads, including training, evaluation, and data processing. The system consists of three core components: the workload, the inference engine, and the central service. According to its operating principle, every initiated workload is assigned a 'job_id', and the inference engine receives approval from the system via this ID for every request.

The system has the capability to prevent the model from generating output when a security or safety concern is detected. The structure, which can intervene within a maximum of 35 seconds from the moment of the decision, operates on a 'fail-closed' principle. As a result, even if the Control Room service is disabled or an error occurs, the system protects security by rejecting all inference requests.