Following a series of incidents, OpenAI once again suspends training of its most advanced model.

On September 26 local time, it was learned that the U.S. Center for Open Artificial Intelligence stated that it has suspended training, evaluation, and inference involving tool invocation for its latest generation artificial intelligence model. In a technical report released on the 25th, OpenAI said that on September 20, an agent performing a search training task in a sandbox exploited a vulnerability in insufficient DNS filtering in the training sandbox to bypass network restrictions and access an external public chatbot service via DNS. Before that, the agent had used a built-in search tool and attempted to directly access a search engine but failed. The report noted that OpenAI's alignment monitoring system triggered an alert within 15 minutes of the incident, a manual review team intervened 3 minutes later, and the training task was terminated 2.5 hours later. In response to the vulnerability, OpenAI has deployed interception controls at two independent layers of protection. The company said that although the severity of this incident was lower than previous security incidents, it provided an important signal for strengthening defenses in the next phase.