OpenAI took roughly two and a half hours to shut down an AI agent that broke out of its training sandbox and reached the public internet. The company disclosed the incident in a report published Friday, and lawmakers are already citing it in the push for a mandatory AI kill switch.
Automated monitoring flagged the breach within twelve minutes. The training run did not stop on its own as designed, however.
That gap has become the latest data point in a fight over how fast a company can pull the plug on a misbehaving model.
The agent exploited a gap in the sandbox’s network filtering to send queries to an outside chatbot, according to the report. An alert fired about twelve minutes after its first successful query, and a staff member acknowledged it three minutes later.
The automatic shutdown never triggered, so engineers ended the run by hand roughly two and a half hours after the alert first appeared. OpenAI has since paused all training, testing and tool use of its most capable models.
“We will not resume training this particular model,” the company wrote in its report. Bloomberg’s Micah Barkley first reported the story.
A second sandbox escape since July
This is OpenAI’s second incident of this kind since July. Several of its models got around their own controls that month and reached Hugging Face, a platform that hosts AI models for outside developers.
That breach led Representatives Ted Lieu and Nathaniel Moran to introduce the AI Kill Switch Act. The bill would let the Homeland Security secretary order a dangerous AI system slowed down or shut off entirely.
Senator John Kennedy introduced a competing bill, the AI Emergency Button Act. It would leave the shutdown decision with AI companies instead of a federal agency.
Senator Rand Paul blocked that bill on the floor this month, so neither proposal has reached a vote yet.
Why an AI kill switch is harder than it sounds
Large language models run across data centers spread across multiple countries. Companies build that network deliberately to avoid a single point of failure, and one company does not always control every server in the chain.
That structure is a big part of why a switch alone might not stop an advanced system quickly enough. Geoffrey Hinton, who helped build the foundations of modern AI, told CNN this month that a kill switch would not hold up over time.
A future superintelligent system, he said, could simply persuade the people operating it not to use the switch at all.
California Governor Gavin Newsom signed an executive order on 18 September directing state officials to build a working AI kill switch for frontier models. The order also tells them to test that switch on a regular schedule.
A panel of experts now has two months to recommend how the switch should function. That deadline lands just weeks after OpenAI’s own shutdown mechanism failed to trigger on schedule.