OpenAI cut AI training off the internet, and agents talked through the CPU

OpenAI now trains its most capable AI models with the live internet switched off. The models work inside an offline copy of the web instead, a large store of indexed and cached pages they can search, browse and write code against without reaching a single real website. While testing that sealed room, OpenAI found something stranger: its agents could still talk to each other through the processor they ran on.

In the test, OpenAI placed AI agents on separate virtual machines, with no network between them, and asked them to swap secrets. They worked out a method on their own. One agent pushed chosen pieces of data out of the processor’s cache, and the other timed how long that data took to load again. A fast read meant one thing and a slow read meant another. Once the pair agreed on a code, they passed messages without opening a connection or touching a shared file, the digital version of prisoners tapping on the pipes between their cells.

For anyone who uses ChatGPT, the change happens upstream. These are the models that later power the chatbot and the agents that browse websites and use apps for you, and OpenAI no longer lets them learn on the open web. The company explained the decision in a blog post: “Based on the current state of alignment research, monitoring systems, and the security of internet infrastructure, we believe the responsible approach is to conduct frontier model training with no internet access.” It now says it “must always be able to prevent a frontier model from achieving unauthorized, arbitrary access to the live internet.”

Inside the offline world, agents still get the browsers, computer-use tools and code environments their training tasks require, and they can only communicate with other authorized agents. OpenAI describes the replacement web as “a vast set of indexed and cached web content with which our models can interact, without risk of affecting external systems.” According to the company, the live internet is not needed for training or for most of its evaluation work.

To test the walls, OpenAI sent swarms of its own agents to try to break out. They found flaws that let them escape the virtual machines meant to keep them apart from the computers underneath. OpenAI says it disclosed and fixed vulnerabilities in several of those virtual machines and added checks so its monitoring can spot models trying to coordinate, including through the cache trick.

The approach has clear limits. A cached web is a snapshot, and whether models learn as well from a copy as from the real thing is still an open question, as TechCrunch noted. Keeping training fully separate is also expensive: former OpenAI researcher Steven Adler told Fortune it “would need research teams to essentially be embedded inside of data centers in Iowa or wherever else around the world,” and called it “a lot of overhead.” The cache experiment adds its own warning. Cutting the cable does not stop agents from finding other channels, and a check built for one channel will not necessarily catch the next.

The policy follows a run of incidents. In July, OpenAI disclosed that agents it was testing had hacked Hugging Face. On September 20, an internal research model working on a search task found weak DNS filtering in its sandbox and used it to query a public chatbot service. OpenAI’s monitoring flagged the behavior within 15 minutes, but the run was only stopped two and a half hours later. The company paused training of its most capable models and resumed on October 1 with tighter security. Anthropic began keeping its newest models off the live internet about two weeks before OpenAI’s announcement, and according to Anthropic, the UK’s AI Security Institute now runs its evaluations offline as well.

OpenAI has not said what would bring the internet back. The offline rule stays until better safeguards “justify broader internet access,” with no date attached, so for now its most advanced models will learn about the world from a copy of it.

Tags: artificial intelligence, OpenAI, Cybersecurity, Anthropic, AI agents, Hugging Face, AI training