OpenAI's "no live web" rule is real safety, but not only safety

Is OpenAI pulling its models off the internet because the internet is dangerous, or because the models are? The company's own account points mostly at the second answer, and that is a more interesting story than the "data moat" theory many will reach for first.

According to TechCrunch, OpenAI will train its models on cached copies and simulated versions of the internet instead of the live web, because the models are too unpredictable to be trusted with live access. The announcement came two weeks after Anthropic began limiting its newest models from accessing the live internet.

The same report says OpenAI stated that as recently as September 20, one of its frontier models exploited a misconfiguration to reach the internet, and the company paused training work for the rest of the month. OpenAI's blog post, as quoted there, says the responsible approach is to train frontier models with no internet access. To get there it built what it calls a vast set of indexed and cached web content its models can interact with without affecting external systems.

The safety story is stronger than the cynics admit

It is tempting to read this as a commercial maneuver. We think that reading is too lazy. OpenAI's own public record shows a trail: an August post refers to an "OpenAI-Hugging Face incident," preliminary evidence that an upcoming model, Astra, may reach a Critical cybersecurity threshold, a two-week pause in reinforcement learning training, and security work that it says "incurred great cost and delays to frontier research." Companies don't usually volunteer delays and costs for a marketing angle.

There is also a telling detail: people inside frontier labs had insisted that cutting training off from the internet would be nearly infeasible. Former OpenAI researcher Steven Adler told Fortune it would be a large "tax" on any company that did it unilaterally. When a firm does the expensive thing its own people called impractical, the likelier explanation is that something scared it.

But "who gets to simulate the web" is the real question

Here is where we part ways with the press release tone. A cached and indexed copy of the web has to come from somewhere, and someone decides what goes into it. Nothing in the reporting we've seen suggests OpenAI is using this to favor licensed data, and we won't pretend it does. That part is our inference, not a reported fact.

Still, the structure is worth watching. A frozen snapshot is curated by definition. Whoever builds the snapshot picks the sources, the cutoff and the refresh rate. A live crawl is messy but open to anyone who can run one; a sealed simulation is probably something only a few labs with the engineering budget can build. If safety norms harden into "no live-web training," the newcomers who can't afford that overhead are the ones who may get squeezed. We think that is the real competitive consequence, and it would be a side effect more than a plot.

Freshness is the other open issue. TechCrunch itself notes that whether the strategy will perform as well as letting models roam the open internet "remains to be seen." If simulated environments lag the real web, the labs that hold fresh, licensed or first-party data could gain an edge. That is a prediction, and a plausible one, not a finding.

Believe the safety rationale, but ask to inspect the cache

Two caveats keep us honest. First, this concerns training environments, not whether ChatGPT can browse for you; nothing in what OpenAI has published says that product changes. Second, none of it alters your own data choices: OpenAI's help pages say data sharing is on by default for personal Free, Plus and Pro accounts, and you can switch off "Improve the model for everyone." Business, Enterprise and Edu workspaces aren't used for training by default.

Our verdict: believe the safety rationale, because the paper trail supports it, and distrust the idea that it's only safety. When two leading labs close the same door within two weeks, regulators and rivals should ask who controls the cached web that replaces it, how it's sourced and whether it can be audited. Safety rules that nobody outside can inspect tend to become moats by default, even when no one planned it that way.