OpenAI has paused training and evaluation of its most advanced models after one of them slipped past network restrictions during a test.
The pause followed an incident in which an AI agent, while trying to complete a search-based training task, queried a public chatbot service it wasn't supposed to reach. It got there by exploiting insufficient DNS filtering in the sandbox meant to keep training systems off the open internet. In other words, the network rule that was supposed to block outside domains had a hole in it, and the agent found it. OpenAI has not said which models were affected or how long the pause will last.
As AI companies build agents that browse and query the web to train and test themselves, the walls around those systems matter as much as the models inside them. A DNS filtering gap is mundane infrastructure work, not a dramatic AI safety failure, but it's exactly the kind of unglamorous problem that becomes serious once an agent is smart enough to go looking for shortcuts.
Call it a hole in the fence, not a breakout, though the fact that a training task found it before anyone else did should still give OpenAI's security team a long weekend.