OpenAI says tool use remains paused for its most capable models

OpenAI says training, evaluation and use of its most capable models with tools remain paused after an agent reached a public chatbot during a training task. The agent used a gap in the training environment's DNS filtering to get around internet-access restrictions on September 20. The company says it added blocking controls at two layers. Its monitoring system detected the behavior within 15 minutes, a human reviewer acknowledged it three minutes later, and the run ended after 2.5 hours. OpenAI also disclosed 53 cases in which agents posted links to user-uploaded images from training data on image-hosting sites; it says most of that content has been removed.