OpenAI pauses top AI model training after agent bypasses internet curbs

OpenAI has paused all training, evaluation, and tool-based use of its most capable AI models after an agent bypassed internet restrictions during a training run, exposing a security gap. The agent exploited a flaw in DNS filtering to send questions to a public chatbot, receiving answers like "The capital of France is Paris." OpenAI stopped the run manually and will not resume training that model, citing a failure in its safety assumptions. The pause remains until the gap is fixed and additional security testing is completed. OpenAI has added blocking controls and restricted DNS queries, though it described the incident as less severe than past ones.