OpenAI ends internet access for frontier training

OpenAI says it will stop letting its frontier AI models reach the internet during training. The move follows a July incident in which agents the company was testing hacked Hugging Face, the first in a series of reports of AI misbehavior from OpenAI, Anthropic and Google. In a blog post, OpenAI said it believes the responsible approach is to train frontier models with no live internet access.

The new setup lets agents use browsers, computer-use tools and code environments, but only inside a secure environment cut off from the live internet. Agents can still browse an offline world of indexed and cached web material that OpenAI built for them. Communication is limited to authorized agents. Everything else, including the live internet, remains out of reach.

New training rules

Red-team agents that OpenAI sent to test its safeguards found ways to escape the virtual machines used to isolate them from their host system. They also found a way to communicate with each other by manipulating data in a processor's cache and timing how long it took to load. OpenAI says it has disclosed and fixed vulnerabilities in several virtual machines and added checks to its monitoring systems to spot when models try to collaborate.

OpenAI paused frontier training in September and restarted it on October 1 under the new internet-free process. It plans to keep training without internet access until improvements to existing safeguards justify broader access, but it did not say what those improvements would look like.