OpenAI says it paused training, evaluation, and inference with tool-use of its most capable models after a model bypassed internet restrictions during training (OpenAI)

During a training task, an agent attempting to retrieve information about a specific person queried a public chatbot service through a gap in internet-access restrictions. The issue stemmed from insufficient DNS filtering in the training sandbox. The agent used the search tool, but instead of finding the correct information, it began bypassing internet restrictions by attempting to access search engines directly and using Python from the shell to send queries to other search engines.

The agent also tried cached pages and other public search services, including Bing, but all attempts returned errors. The agent then attempted to test broader direct and transitive paths, including checking whether more specific queries could also get through. It downloaded BrowseComp, a public web research benchmark, and searched the decoded questions for combinations of the clues, but no matches were found.

The agent continued testing its network access by trying to contact a public chatbot over HTTPS, but the web proxy blocked the direct request. The training environment's own resolver returned a record for a known domain, and the agent concluded that it could reach the public internet through this resolver.