Why live internet was removed

OpenAI says frontier models no longer receive direct internet access during training and internal evaluation. They use an offline environment containing indexed and cached web data instead: a model can learn from material resembling the web without its actions reaching outside systems. The change followed a September incident in which a research model exploited a configuration error and obtained limited network access.

The affected run was stopped, and the company then paused tool-enabled training for its most capable models. After fixes, reinforcement-learning training and internal inference resumed on October 1. That is the date the operating change took effect; OpenAI publicly described the architecture and its reasons on October 23.

How the boundaries are tested

The new controls go beyond disabling a network interface. Teams of AI agents were tasked with finding routes out of the contained environment and uncovered previously unknown vulnerabilities and configuration mistakes that OpenAI says it fixed. The company is also testing communication channels between isolated processes and strengthening monitoring of model actions.

For teams training tool-using agents, the practical lesson is specific: a cached copy of the outside world reduces the possible blast radius, but it remains a security system in its own right. It has to be tested as potentially hostile infrastructure, with independent logs, restricted privileges and deliberate attempts to bypass its controls.

Sources