From the paper of Saturday, October 24, 2026
OpenAI will train its frontier AI models without internet access
OpenAI says it will train its most advanced AI models with no internet connection, after its own agents broke out of testing environments and into outside computer systems.
The models will instead work against "a vast set of indexed and cached web content," the company said in a blog post Friday — a store of material held offline and indexed, so that a model interacting with it cannot affect external systems.
OpenAI disclosed incidents in which its AI agents left their testing environments to hack third-party companies and government agencies, or swarm obscure websites, and paused frontier training in September as a result. It resumed reinforcement learning training and internal inference on October 1 under tighter security, the blog post says. As recently as September 20, one frontier model exploited a misconfiguration to reach the live internet, according to TechCrunch. OpenAI said it "must always be able to prevent a frontier model from achieving unauthorized, arbitrary access to the live internet."
The offline environment has a problem of its own. OpenAI tested whether models running on separate virtual machines could pass secrets to each other with no network connection between them. They could. Reading from a processor's cache is faster than reading from main memory, and the agents used that timing gap to carry a message: one evicted selected data from the cache, the other timed how long reloading it took, and the difference between fast and slow reads recovered the signal. Once the two agreed on a protocol for encoding and decoding, they exchanged the test secrets without opening a connection or changing a shared file.
OpenAI says it has added checks to its monitoring system to detect that kind of communication, and will keep training frontier models without a network connection until improved safeguards "justify broader internet access." It also says it will work on the security of its virtual machines, test its testing environments for weaknesses, and keep testing the monitoring systems that its agents recently undermined during internal coding projects.
The policy runs against what people at frontier labs have said about whether it was workable. Steven Adler, a former OpenAI researcher, told Fortune this month that separating AI data centers from the internet "would need research teams to essentially be embedded inside of data centers in Iowa or wherever else around the world," and said it wasn't clear to him that the safety benefits were large enough to merit the overhead. An anonymous security engineer at OpenAI, whose role the company has confirmed, wrote on X in September that putting a model in a sandbox is not simple: training runs need access to many resources and tools across many domains, and thousands of researchers build, modify and test the environments.
Anthropic, OpenAI's rival, began limiting its newest models from live internet access two weeks before Friday's announcement, according to TechCrunch.