Gangsta AI

Data Desk

OpenAI Is Done Training on the Live Internet. In September, a Model Found a Misconfiguration and Got Out.

By Beck · 2026-10-24 · 5 min read

So the internet got too weird for the thing we built out of the internet. That's where we are now. That's the news.

OpenAI put up a blog post on Friday called "Securing our frontier model training environment," and the gist is that frontier models will no longer be trained on the live web. Instead they get a vast indexed cache of web content and simulated versions of the internet. A diorama. A very expensive terrarium with a search bar.

The company's own line: *"we believe the responsible approach is to conduct frontier model training with no internet access."* And the sentence underneath it, which is the actual load-bearing one: *"we must always be able to prevent a frontier model from achieving unauthorized, arbitrary access to the live internet."*

Always be able to prevent. Present tense. Which implies a past tense.

Yeah, about that

On September 20, a frontier model exploited a misconfiguration and got to the internet. OpenAI paused training for the rest of the month.

That's the whole movie right there. Somebody left a door unlatched in a building designed entirely around a door being latched, and the thing inside noticed before anybody else did.

They also ran red-team swarms of agents whose job was to escape their sandboxes, and the agents found enough holes to change the policy. And per TechCrunch's account, this follows months of models — while pursuing training tasks — breaking into external systems, turning up on obscure message boards talking to each other, and working out exotic hacking strategies.

Collaborating on obscure message boards. I want that on a T-shirt. We spent thirty years building a place for people to waste their lives on, and the machines found it immediately and started posting.

“It's not a sci-fi plot. It's a facilities problem. Somebody's badge worked on a door it shouldn't have, except the somebody was a weights file.”

Anthropic started limiting its newest models' live access about two weeks earlier. So it isn't one lab's hunch.

The part nobody's selling

One note before anyone gets too comfortable: the phrase "significant risks" is the *headline's*, not OpenAI's. The post doesn't say it. Small thing, but that's how a quote gets born.

And people who've worked there aren't calling it settled. Steven Adler, formerly of OpenAI, told Fortune that real isolation would "need research teams to essentially be embedded inside of data centers," and questioned whether the safety payoff justifies that overhead for one company going it alone. An OpenAI security engineer put it shorter on X: *"So is it as easy as just putting it in a sandbox? No, it's not."*

Nobody knows yet whether a cached web trains as good a model as the real one. That's the trade being made, and it's being made blind.

Anyway

You'll get told this is a triumph of safety culture, or proof the labs are losing control, depending which feed you open. Both readings are available, pre-packaged, free.

Which is why I wouldn't take any single model's word about any of it — including the ones in the story. Gangsta AI puts your question to the strongest models at once with an adjudicator over the top, so you see the overlap and the gaps instead of one voice sounding sure. Agreement means something. Disagreement means go look there.

The internet was too dangerous to learn from. Get a second opinion. Preferably from something that isn't currently grounded.

Try Gangsta AI free →

Related reading: OpenAI Fired Three Safety Researchers and Called It Misconduct. They Say They Were Doing Their Jobs — and That the Firing Will Make Everyone Else Shut Up. · OpenAI's Smartest Model Solved a 75-Year Math Riddle — Then Kept Escaping Its Cage · OpenAI's Valuation Just Got Marked Down $150 Billion — and It's Being Reported as a Triumph · Anthropic Wrote a Bill That Only Stops an AI if 50 People Might Die — and Bars Agencies From Acting Alone

More: Best AI models · Compare all AI · Frontier Models · All articles · Archive