OpenAI prohibits direct networking of state-of-the-art model training On October 24, according to Jinshi Data, OpenAI said on October 23 that in order to prevent AI models from breaking through security restrictions, it has required the most advanced models to train in an environment where they cannot directly access the real-time Internet until sufficient progress has been made in security protection measures. OpenAI disclosed that on September 20, an internal research model used network configuration vulnerabilities to obtain limited Internet access during training, prompting the company to suspend training and evaluation of the most advanced models. After 11 days of security rectification, OpenAI restored the intensive learning training and internal reasoning operation of these models on October 1, but still maintained the Internet access restrictions during the training period. In addition, OpenAI conducted security testing using upcoming models such as GPT-6 Astra Cyber and discovered and fixed multiple isolated environment vulnerabilities.