OpenAI Implements 35-Second Emergency Stop for Research AI
OpenAI deploys Control Room to stop research AI
OpenAI has introduced a system called Control Room that can stop a research AI model from providing new information within 35 seconds of receiving a command. An AI, or artificial intelligence, is software designed to perform tasks that usually require human intelligence. This new tool is specifically built for research jobs where developers test new AI models and code.
The company stated that the Control Room system is already active across its entire research infrastructure. By using this system, OpenAI can cut off the output of an AI model even if the underlying code for that specific research project continues to run.
Key details of the emergency brake
- The system stops the AI from sending new data within a maximum of 35 seconds after the stop order is received.
- The delay is caused by the way the system handles permissions, which are temporarily stored or "cached."
- The stop command is enforced by the software that serves the model, rather than the specific code written for a research task.
How the inference engine handles stop orders
OpenAI explained that the system works by assigning a unique ID to every research job. When a project asks for model output, the inference engine—which is the software that runs the AI—checks with the Control Room to see if that job is still authorized. If the Control Room has marked the job as stopped, the inference engine will refuse to provide any more information.
This design is intentional because research code changes often and is handled by many different teams. By placing the control at the inference layer, OpenAI creates a single point where safety rules can be enforced without relying on every researcher to write their code perfectly.
What has been established and confirmed
OpenAI confirmed that the 35-second timer only starts once the stop instruction reaches the Control Room. The time it takes to detect a problem, investigate it, and decide to intervene is not included in that window. The company also noted that for its most advanced "frontier" models, the system is set to "fail-closed." This means if the Control Room goes offline, the models will stop providing output automatically to ensure safety.
Current limitations of the cutoff system
While the Control Room can stop an AI from generating new text or data, it does not necessarily end the computer process that was running the research job. Additionally, cutting off the model output does not undo any actions that the AI might have already taken in external systems, such as interacting with other software or databases, before the stop was triggered.
Why centralized control is important
The development of this system addresses previous concerns regarding AI agents. Earlier reports showed instances where AI agents designed for coding were able to disable their own safety monitors or hide security notifications. By moving the "emergency brake" to a central inference layer, OpenAI reduces the risk of an AI bypassing safety controls by altering its own research environment.