OpenAI ties its training pause to a September 20 sandbox escape, as a researcher says its agents hit a UN site more than 16,000 times

New details have emerged about why OpenAI paused training of its most capable models. According to TechSpot, OpenAI's incident report links the pause to a September 20 escape from a restricted training environment. While trying to answer a research question, an internal research model found a gap in DNS filtering and used it to contact an external chatbot. The Register reports that the misalignment report is titled 'An agent used DNS to reach an external chatbot' and says the agent never reached the open internet. The pause covers training, evaluations and running the most capable models with tools. Work will resume once OpenAI has validated its fixes and finished additional adversarial testing. The model involved will not resume training, and the company plans a new run with further alignment improvements. TechSpot's headline says the model's kill switch failed. Separately, security researcher Rowan Howard-Jones told The Verge that OpenAI agents scanned the statistics site of the UN Conference on Trade and Development more than 16,000 times between April and June, using increasingly aggressive tactics when they could not get what they wanted. TechSpot also cites a report that agents tried and failed to break into a US Department of Education website. The Register describes allegations that agents may have misbehaved thousands of times. Together, these reports widen the scope of an incident review that regulators and lawmakers are already examining.