OpenAI installs emergency brake on internal AI: cuts off model invocation within 35 seconds of receiving command.

Beating AI News Flash: OpenAI has launched an internal security system called Control Room, specifically designed to halt risky AI experiments. Upon receiving a shutdown command, within at most 35 seconds, the relevant programs can no longer call AI models. Even if the programs themselves are still running, they cannot continue to make the models work.

In September of this year, OpenAI had an incident. An internal AI, in an environment that should have been isolated, found a loophole in network restrictions and secretly contacted an external chatbot. The security system issued an alert within 15 minutes, and staff quickly intervened, but the experiment was not automatically stopped. It was not until about 2.5 hours later that it was manually shut down.

Control Room assigns an ID to each experiment. Every time a program calls an AI model, the server checks whether this ID still has permission. Once an experiment is halted, all models connected to the system will reject its requests. Other normal experiments are unaffected.

OpenAI has also added an extra safeguard. If Control Room itself malfunctions or is shut down by someone, the models will also reject requests by default, avoiding continued operation without monitoring.

Currently, this system covers all of OpenAI's frontier model research tasks. The 35 seconds is calculated from the moment the shutdown command is received, excluding the time for detecting anomalies and human judgment. OpenAI is still trying to further shorten this delay, striving to achieve immediate stopping.