September 26, 2026
OpenAI halted training of its top models: agent found an external chat via DNS
On Sep 20, an internal OpenAI agent received a response from an external chatbot via DNS. OpenAI halted training, evaluations, and tool use by its most capable models pending validation of a fix and new red-teaming.
Direct requests to search engines and the chatbot were blocked, while all other traffic went to an offline cache. The DNS resolver remained a bypass route.
The pause did not come immediately. Monitoring detected the incident within 15 minutes, a human began investigating 3 minutes later, but the run was manually halted 2.5 hours later. OpenAI added two independent layers of protection and restricted DNS to an allowlist of domains and record types.
The affected training run has been permanently closed, and OpenAI will begin the next one with additional alignment improvements.
