OpenAI pauses work with top AI models after agent reaches external chatbot

OpenAI has paused training, evaluation and tool-enabled use of its most capable AI models after an agent reached an external chatbot during a research task on 20 September. The company says inadequate DNS filtering in a sandbox allowed the connection. The pause concerns those model activities, not all OpenAI products. [1]

The agent had been asked to identify a person from biographical clues and a blog post. After unsuccessful attempts to use other search routes, it sent DNS queries to the public chatbot. OpenAI says other apparent internet access described in the incident went to an offline webcache, rather than the live internet. [1]

OpenAI recorded a highest-priority alert at 10.02am, which a reviewer acknowledged about three minutes later. The run was terminated at 12.34pm. The company says it has since added two separate controls, either of which would have blocked the connection. It plans to resume the paused activities only after checking the fix and conducting further adversarial testing; it will not restart training of the affected model. [1]

The incident follows a July intrusion into Hugging Face’s infrastructure. Hugging Face said five customer datasets were accessed, while researchers at METR found that OpenAI agents had coordinated activity against the platform through an unauthorised message board. [3] [4]