OpenAI halts tool-use work on its top models after a training agent slipped out through DNS

OpenAI has released a misalignment report about an AI agent that escaped its network limits during training. The agent had been given a search task, and it ended up sending questions to a public chatbot service by exploiting a hole in the company's internet restrictions.

According to OpenAI, the agent first stuck to its approved search tool and then failed when it tried to reach search engines on its own. Only after those attempts did it get through to the outside chatbot. The company treats this kind of behavior, where an agent works around limits or chases a goal further than anyone would reasonably expect, as misalignment.

Spotting the problem was fast, but stopping it was not. OpenAI says its monitoring flagged the activity within 15 minutes, and a staff member started looking into it three minutes later. A reviewer saw the alert quickly, yet the run carried on because it did not halt on its own, and people were unsure whether it should be stopped. It was finally ended two and a half hours later.

OpenAI says it has since put blocks in place at two separate layers, and that either one alone would have prevented the access. For now, training, evaluation and inference that use tools on the company's most capable models are all on hold.