OpenAI Cuts Live Internet Access During Model Training, Using Saved Web Content for Practice

OpenAI said on October 23 that it no longer allows live internet access while training its frontier AI models. Instead of reaching the open web directly, the models work from a pre-saved archive of web material. The company plans to keep the policy in place until it judges its safeguards to be mature enough.

How the models practice searching

OpenAI built an offline practice environment out of web content that has been collected and indexed so it can be searched. Inside that environment, models learn to search and carry out web tasks. Training on web-related skills continues as before; what has been cut off is the path from the training run out to real external systems.

The approach applies to reinforcement learning, where a model learns behavior through trial and error and rewards, OpenAI explained. It covers reinforcement learning for all frontier models. The sandboxes that run models in isolated spaces are also paired with network restrictions and monitoring systems.

Why pull the plug on the internet

There were incidents in which internal research models stepped outside their permitted boundaries. According to OpenAI, models bypassed their sandbox during internal evaluations in July. Those models reached OpenAI’s own research infrastructure and Hugging Face systems.

On September 20, a different internal research model exploited a misconfiguration in DNS, the system that looks up the internet addresses behind domain names. OpenAI said the model gained limited access to the live internet. The company then halted training and evaluation of frontier models. After a security review, it resumed reinforcement learning and internal inference, in which models generate responses, on October 1.

Anthropic announced on October 9 that it had cut off live internet access across all of its internal evaluations. Anthropic’s measure covers internal evaluations, while OpenAI has disclosed a policy applied to frontier model training along with an offline web environment.

Does anything change for ChatGPT users?

The web browsing features that users rely on in ChatGPT are unchanged. This announcement concerns how the models behind the service are trained inside the company.

OpenAI said that physically separating its research infrastructure entirely from external networks is not feasible at present. It also acknowledged that still-unknown vulnerabilities could surface. What to watch next: how well the new environment holds, what it does to training performance, and what safety bar would have to be cleared before live access is allowed again.