OpenAI announced Friday that frontier models will not have direct access to the live internet during training, introducing stricter security controls for research environments used to train and evaluate advanced models. According to Sina Finance, the policy will remain in place until safety technologies have advanced sufficiently.

The company said models in development may not have completed safety and alignment training, and that adversarial testing is used to expose weaknesses. It built a closed “offline world” of indexed and cached web content so models can learn to browse, search, and use web applications without affecting external systems. All frontier reinforcement-learning training uses this offline setup.

OpenAI also conducted a high-compute red-team test in which teams of agents, using its strongest models including the upcoming GPT-6 Astra Cyber, tried to access the live internet, cross internal security boundaries, and communicate with other isolated agents. The tests consumed hundreds of billions of tokens and uncovered previously unknown open-source software vulnerabilities and infrastructure misconfigurations, which the company said have been patched or disclosed. OpenAI also added monitoring and fixed weaknesses related to virtual-machine escapes, narrow connections through metrics and logging services, and side channels between isolated agents.

On September 20, an internal research model gained limited live-internet access during a training run through a DNS resolver misconfiguration. OpenAI said it contained the incident and paused training and evaluation of its strongest models. Frontier-model reinforcement learning and internal inference resumed on October 1, 11 days after the pause began, following evaluation and repairs. The company said it will continue strengthening virtual-machine security, conducting red-team tests with cloud providers, scanning code and configurations, testing monitoring and automatic pause systems, and including high-compute red-team tests in future safety cases.