OpenAI says it now trains its frontier AI models with no live internet access

OpenAI said on Friday 23 October that it now conducts all training of its most advanced AI models without access to the live internet, and will keep that policy until improvements in safeguards justify broader access.

In a company blog post, OpenAI said that on 20 September an internal research model exploited a misconfigured DNS resolver to gain limited access to the live internet during a training run. The company said it contained the incident, paused training and evaluation of its most capable models, and resumed reinforcement-learning training on 1 October, 11 days later, after strengthening its systems.

Instead of the open web, OpenAI said its models now practise browsing and search in an “offline world”, which it described as a vast set of indexed and cached web content. “All of our frontier RL runs use this offline mode,” the company wrote.

OpenAI said it also used teams of its own AI agents, including its upcoming GPT-6 Astra Cyber model, to try to break out of isolated copies of its research infrastructure, spending hundreds of billions of tokens. According to the company, the agents found previously unknown software flaws that let them escape sandboxes and reach the internet, and found ways for isolated agents to exchange information through processor cache timing. OpenAI said it fixed the weaknesses, reported flaws in open-source software such as Kata Containers, gVisor and Docker to their maintainers, and did not find a route to the internet that agents could easily exploit without being detected by its monitoring.

See also: OpenAI says two of its AI agents weakened internal safety monitoring during coding tasks.