Stopped in 35 secs: How OpenAI's Control Room will monitor AI safety risk
Control Room works alongside OpenAI's misalignment monitoring systems to identify potential issues and stop a job when intervention is required
Control Room is already deployed across OpenAI's research infrastructure, where the company said the need for such monitoring is greatest. (Photo: Reuters)
Listen to This Article
OpenAI earlier this week introduced a central monitoring system called Control Room to track workloads across its research infrastructure, detect potential issues, and quickly halt model outputs within 35 seconds of a decision to intervene. This comes amid growing concerns over the safety and security risks posed by increasingly powerful artificial intelligence (AI) models.
Announced on Monday, the system connects research inference workloads, including agentic sampling for model training and evaluation, data processing, and other research jobs. It works alongside OpenAI's misalignment monitoring systems to identify potential issues and stop a job when intervention is required.
Control Room has been designed to remain operational even if other systems fail. It also follows a ‘fail-closed’ approach, under which all inference requests are denied if the Control Room service itself becomes unavailable, the ChatGPT maker said.
The system is already deployed across OpenAI's research infrastructure, where the company said the need for such monitoring is greatest. Its external model-serving infrastructure is more centralised and already has operational systems to take down model deployments when required.
How Control Room works
OpenAI's Control Room integrates three components: the research workload, the inference engine and the central Control Room service.
Also Read
A workload refers to a research job, such as a reinforcement learning experiment, an evaluation or a data-processing task. Every research workload involving agentic inference must be linked to an approved safety case, OpenAI said.
When a researcher starts a job, the workload code requests a unique job ID from Control Room. Whenever the job needs to generate output from a model, it sends that ID to the inference engine, which checks with Control Room whether the job is authorised to continue.
If a monitoring system detects a potential issue and decides that a job should be stopped, it sends a termination request to Control Room using the job ID. The system records the job as terminated and denies subsequent inference requests, preventing the model from generating further output for that job.
System designed to remain operational during failures
OpenAI said the maximum time between receiving an instruction to stop a job and preventing further model output is 35 seconds. This delay is caused by a caching layer used to handle the volume of inference requests. The company said it is working to reduce the delay to zero.
The system's termination mechanism operates at the inference layer rather than relying on individual workload code. OpenAI said research workloads are decentralised and frequently change, making them unsuitable for hosting critical safety protections.
The inference layer, by contrast, has a smaller number of centralised and uniform code paths. This allows Control Room to check authorisation for every inference request across research workloads.
The company said Control Room is one part of its broader approach to ensuring the safety and security of frontier model training and deployment. It plans to expand the system alongside investments in monitoring for misaligned behaviour and security breaches.
More From This Section
Don't miss the most important news and views of the day. Get them on our Telegram channel
First Published: Nov 12 2026 | 11:22 AM IST
