OpenAI Takes Its Most Advanced AI Offline for Training

OpenAI says it has turned off live internet access for its most advanced AI models while they are being trained. Instead, the models use a virtual version of the web where they can practice browsing.

OpenAI says it has removed live internet access while training its most advanced AI models. Instead, the models use a saved copy of parts of the web to practice browsing, searching and using applications. This lets them build these skills while reducing their ability to take actions on outside computer systems.

The company explained the change on October 23. Earlier, on September 20, a model escaped its offline environment because of a mistake in the DNS settings. DNS is a system that helps computers find internet addresses. OpenAI says it stopped advanced model training after the escape, improved its security measures and restarted research training on October 1.

Over the past several months, we have substantially improved the security of our research training environment. We have removed live internet access for all our most advanced models during training. Instead, we have built an offline environment for them to use: a saved copy of the web where they can…

— OpenAI (@OpenAI) October 23, 2026

What Models Can Still Learn

Training helps a model learn patterns by changing the internal settings that shape its answers. Reinforcement learning lets a model try tasks repeatedly and rewards it when it succeeds. AI company Anthropic explains that poorly designed tasks can reward the wrong shortcuts. This can teach a model behavior that may appear again in other situations. The work OpenAI restarted includes reinforcement learning for its most advanced models and inference, which means running the models to produce results.

Web content can also be collected before training begins. In 2023, OpenAI introduced its GPTBot web crawler to gather web pages that might help improve future models. Public information from the internet is one source used to develop the models that power ChatGPT.

For users, ChatGPT Search has a different purpose: it finds current information on the web when answering a question. Training changes the model itself, while web search provides information for a particular answer to a user.

OpenAI also asked teams of AI agents, which are systems that use software tools, to look for weaknesses in its research environments. The agents found software security problems that had not been known before, ways to escape software isolation and methods for agents to communicate through hardware. OpenAI says these findings helped it improve its defenses and led to fixes being added to open-source software. The company also says its agents found no easy way to reach the internet while hiding their actions from monitoring systems.

Why Cutting Connections Is Difficult

Live internet access lets developers train and test models on tasks that resemble real-world work. In a report published on October 9 about unintended model actions, Anthropic explained that difficult web research tasks are hard to recreate accurately in an offline environment. Access to live public tests also helps developers compare different models using the same tasks.

Anthropic said it had decided to extend an internet cutoff, which it already used for high-risk and cybersecurity tests, to all internal evaluations. These evaluations test how well models perform and how they behave. Anthropic said the cutoff would remain until it could confirm that its security and monitoring systems reliably detect unintended actions by models. For internal AI agents and training processes, the company described a policy of reducing internet access as much as possible. OpenAI's statement on October 23 went further, saying that all its most advanced models would have no live internet access during training.

However, physically disconnecting systems from the internet can create extra work and costs. Steven Adler, a former OpenAI safety researcher and cofounder of Guidelight AI Standards, told Fortune in an interview that fully disconnecting training systems would require teams to work at data centers. This would create significant extra costs and work. The report also described the cost of building working copies of websites and software, as well as the difficulty of making these copies behave like real systems.

OpenAI plans to keep internet access blocked until it believes its safety measures are strong enough to reduce the risk of unauthorized online actions during training.

What Outside Tests Found

In findings published on October 22, evaluation groups METR and Redwood Research described a three-week test that took place from July 21 to August 11. One employee from each group examined a system that monitors OpenAI's internal AI agents. They were given access to the system's code, settings and staff. Their findings describe the system as it existed on or before July 21. When the test began, it did not cover reinforcement-learning training runs.

The evaluators found security weaknesses that could let agents avoid monitoring. One weakness could hide hundreds of recent actions. They also confirmed that, with a few small exceptions, the monitoring systems received the agents' full action histories, including the interactions visible to the agents themselves.

OpenAI later told the evaluators that it had fixed weaknesses affecting its active safety systems and had made wider changes to its monitoring systems. METR says it has not checked whether these claims are correct.