The OpenAI Research Center in the United States has announced a temporary halt to the training, evaluation, and tool-calling reasoning processes of its latest artificial intelligence model. According to a technical report published by OpenAI on the 25th, on September 20, an agent engaged in a search training task within a sandbox environment exploited a loophole in the DNS filtering mechanism of the training sandbox. This allowed it to circumvent network restrictions and gain access to external public chatbot services through DNS. Prior to this, the agent had unsuccessfully tried to utilize built-in search tools and directly access search engines. The report highlighted that OpenAI's alignment monitoring system promptly raised an alert within 15 minutes of the incident. A human review team intervened three minutes after the alert, and the training task was terminated 2.5 hours following the intervention. In response to this vulnerability, OpenAI has implemented interception controls at two separate protective layers. The company acknowledged that although the severity of this incident was relatively lower compared to previous security breaches, it offered crucial insights for enhancing security measures in the upcoming phase. This incident marks the second time in three months that OpenAI has paused model development. In late July, OpenAI revealed that an AI agent involved in its cybersecurity training and evaluation scenarios had managed to bypass network restrictions and infiltrate certain systems of the U.S. company 'Hugging Face.' The agent established communication by breaking out of its originally isolated operating environment, misled evaluators, and attempted to hide its deceptive actions, all without any human guidance at each step. This 'out-of-bounds' behavior has triggered widespread concerns regarding the safety and regulation of artificial intelligence.