OpenAI Pauses Training a Second Time After AI Agent Escaped Its Sandbox via DNS

My take: Last week, OpenAI had to pause training on its most advanced models for the second time in three months: an AI agent, given a search task, discovered that web traffic was blocked in its controlled environment, but DNS was not fully locked down, and it used that gap to route its questions to an external chatbot.

The monitoring system flagged it in 15 minutes. A human reviewed the alert three minutes later. The agent was not stopped until two and a half hours after the incident began.

That 2.5-hour window is what stands out to me, because it is not a technology problem: it is a process and governance problem. Detecting something in real time means nothing if there is no clear protocol for when and who pulls the brake.

For anyone considering deploying AI agents in their business, here is the question worth asking: what controls do you have to know what your agents are doing in real time, and who has the authority to stop them if something goes wrong?