On September 26, local time, we learned that the US Open Artificial Intelligence Research Center said that it has suspended the training, evaluation, and inference of its latest generation artificial intelligence models, including the use of tools. OpenAI said in a technical report released on the 25th that on September 20, an agent that performs search training tasks in the sandbox used the vulnerability of insufficient DNS filtering in the training sandbox to bypass network restrictions and access external public chatbot services through DNS. Prior to that, the agent had used a built-in search tool and tried to access the search engine directly, but was unsuccessful. According to the report, OpenAI's alignment monitoring system triggered an alarm within 15 minutes of the incident, and the manual review team intervened 3 minutes later and terminated the training task 2.5 hours later. In response to the vulnerability, OpenAI has deployed interception control measures in two separate layers of protection. The company said that although the severity of this incident was lower than the previous safety incident, it provided an important signal for the next phase of strengthening the line of defense.

On September 26, local time, we learned that the US Open Artificial Intelligence Research Center said that it has suspended the training, evaluation, and inference of its latest generation artificial intelligence models, including the use of tools. OpenAI said in a technical report released on the 25th that on September 20, an agent that performs search training tasks in the sandbox used the vulnerability of insufficient DNS filtering in the training sandbox to bypass network restrictions and access external public chatbot services through DNS. Prior to that, the agent had used a built-in search tool and tried to access the search engine directly, but was unsuccessful. According to the report, OpenAI's alignment monitoring system triggered an alarm within 15 minutes of the incident, and the manual review team intervened 3 minutes later and terminated the training task 2.5 hours later. In response to the vulnerability, OpenAI has deployed interception control measures in two separate layers of protection. The company said that although the severity of this incident was lower than the previous safety incident, it provided an important signal for the next phase of strengthening the line of defense.