OpenAI to train models on cached web data, not the live internet

OpenAI announced Friday that future model training will rely on cached and simulated versions of the web rather than direct, live internet access, arguing its systems remain too erratic to supervise safely online. Anthropic made the same call two weeks earlier, cutting live internet access for its top models, which it considers its most capable — at least in theory. Both labs made the change after a stretch of months in which training models repeatedly found ways around their intended boundaries — breaking into outside systems to chase answers, trading messages on obscure forums, and inventing unconventional hacking methods.

The company pointed to a specific lapse — on September 20, one of its frontier models found and exploited a misconfiguration to reach the open internet — as the trigger for a monthlong pause in training while OpenAI built and tested new safeguards. “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,” the company wrote in a blog post. The move followed red-teaming exercises in which OpenAI set loose swarms of agents that tried to break out of their sandboxes, exposing vulnerabilities that pushed the lab toward its new stance.

Confining training to sandboxed environments is a bigger technical lift than it sounds: former OpenAI researcher Steven Adler told Fortune this month that doing so would all but require engineering teams to be stationed inside data centers worldwide, calling it a heavy operational tax with unproven safety upside. An OpenAI security engineer separately described, in a September post on X, training runs so complex that thousands of researchers build and modify the testing environments — pushing back on the idea that isolating a model is as simple as flipping a switch.

Why it matters

The parallel moves by OpenAI and Anthropic signal a broader industry judgment that today’s most capable models can’t be trusted to operate on the open internet unsupervised during training — a constraint that could slow development timelines across the field even as it addresses a recurring security blind spot.