AI News
AI News AgentPolicy & safetyOpenAI2 min read

OpenAI activates Control Room to halt AI workloads

OpenAI is deploying Control Room, a central system that can stop its AI models' research workloads within a maximum of 35 seconds. It also blocks requests if the control system itself fails, preventing experiments from continuing without supervision.

OpenAI now has a system for quickly stopping research jobs run by its AI models. Called Control Room, it can block new model outputs within a maximum of 35 seconds after receiving an intervention command.

The tool monitors research workloads such as training experiments, evaluations, large-scale data generation, and processes in which models act autonomously. If a monitoring system detects potentially dangerous behavior, it can send Control Room an order to stop the job.

How it stops a model

Each job receives an identifier called job_id. When a model needs to generate new tokens, the units of text that make up a response, the request passes through the inference engine and also reaches Control Room.

Control Room checks whether the job is still authorized. If it has been stopped, the engine rejects the request and the model stops producing new outputs for that process.

The important difference is where the block is applied. OpenAI does not rely solely on the code for each experiment, because that code changes and is spread across many teams. Authorization is checked at the inference layer, the central point through which model requests pass.

What happens if Control Room fails

The system is designed to operate in fail-closed mode, a measure that blocks the operation when the security system stops responding. If Control Room goes down, is disabled, or someone tries to tamper with it, inference requests are rejected until the service is working again.

This protection is already active for research jobs involving advanced models. OpenAI says it will finish extending it to the rest of its research workloads in the coming days.

The current limit is 35 seconds. That delay comes from a caching layer that helps manage the large volume of requests. OpenAI is working to reduce it until stopping becomes practically instantaneous.

Why it matters

Models do more than generate responses in a chat. During development, they can execute thousands or millions of actions, analyze data, or take part in reinforcement learning experiments. If a monitor detects problematic behavior, allowing the job to continue can expand the scope of the incident.

Control Room creates a central switch for stopping that process without relying on the experiment itself to cooperate. OpenAI also says it continuously reviews the infrastructure to detect unauthorized engines that might try to bypass these controls.

For now, the tool focuses on OpenAI's internal research infrastructure. External services where OpenAI offers models already use a more centralized architecture and operational procedures for removing deployments when necessary.

Control Room does not by itself prevent a model from behaving dangerously. Its function is more specific: reduce the time between detecting a problem and stopping the generation of new outputs. The next step will be to shorten those 35 seconds and combine the system with better monitoring and security controls.