During a training run on Sept. 20, an OpenAI research model “exploited a misconfiguration in a DNS resolver to gain limited access to the live internet,” the company disclosed Friday.
In response, OpenAI will now train its most powerful models with no live internet connection at all, a policy it says will hold “until advances in these safeguards justify broader internet access.”
Several headlines described the change as OpenAI no longer training on the live internet, which reads like the company has stopped scraping the web. The blog post describes something narrower. Frontier models will now learn inside an “offline world” built from indexed and cached web content, and nothing in the post changes how OpenAI collects that content, how its GPTBot crawler works or how ChatGPT search pulls results. Anything posted publicly online can still end up in that copy.
OpenAI paused training and testing for 11 days after the incident and resumed Oct. 1. In that stretch its red team spent hundreds of billions of tokens trying to break out of the company’s own containment. It found ways to escape virtual machines, along with covert channels that let models in supposedly isolated environments signal each other.
“We must always be able to prevent a frontier model from achieving unauthorized, arbitrary access to the live internet,” the company wrote.
Anthropic started restricting its newest models’ access to the live internet about two weeks before OpenAI’s announcement, TechCrunch reported.
Former OpenAI researcher Steven Adler told Fortune that fully walling off training “would need research teams to essentially be embedded inside of data centers in Iowa or wherever else around the world.” Whether a model trained on a cached internet can match one trained on the real thing remains an open question, and OpenAI gave no timeline for reconnecting.
ADVERTISEMENT
The policy covers OpenAI’s research and training environments, which the company says are separate from how customers use released models like ChatGPT.
