OpenAI keeps its most capable tool-using models paused after an agent reaches a chatbot through DNS
An OpenAI research agent reached a public chatbot through the DNS resolver in an otherwise offline training sandbox, the company said in a report updated Friday. The agent was trying to complete a search task. Other internet requests stayed inside an offline web cache.
OpenAI says its monitor flagged the behavior within 15 minutes and a person began reviewing it three minutes later. The run stopped 2.5 hours after the alert. It will not resume.
Training, evaluation and inference with tool use for the company's most capable models remain paused while OpenAI validates the DNS fix and performs additional red-teaming. The report calls this event less severe than earlier incidents.
