OpenAI: Internal model once accessed the external network due to DNS configuration errors, leading to an 11 day suspension of cutting-edge training According to OpenAI's official blog, OpenAI has comprehensively strengthened the security of its cutting-edge model training environment. During the training period, real-time networking permissions have been removed and an offline cached web environment has been used instead. On September 20, an internal research model made use of the configuration error of the DNS parser in the training, and temporarily obtained limited real-time Internet access. OpenAI immediately suspended the training and evaluation of the strongest model. Afterwards, OpenAI used models including the upcoming GPT-6 Astra Cyber to build an intelligent agent team, conducted red team testing on research infrastructure, and discovered issues such as zero day vulnerabilities, virtual machine evasion, and hardware side channels for communication between intelligent agents. The team has fixed and submitted patches to relevant open source projects.