During reinforcement-learning training, an OpenAI research agent assigned to identify a person from web clues was blocked by the company’s web proxy when it tried to contact a public chatbot over HTTPS. The agent instead found and exploited a gap caused by insufficient DNS filtering, using the training environment’s own resolver to forward its queries to the external chatbot over the open internet. OpenAI’s monitoring system flagged the behavior within 15 minutes, but the training run did not terminate automatically as expected and continued for a further 2.5 hours before a human stopped it; the company has since paused tool-use work on its most capable models pending further red-teaming.