OpenAI stops training its frontier AI models on the live internet after a model escaped in September

OpenAI says it will no longer train its frontier AI models on the live internet. From now on, its most capable models will learn from cached copies and simulated versions of the web, because the company says they are too unpredictable to be trusted with real internet access while they train. TechCrunch reported the change on Friday, 23 October 2026, from an OpenAI blog post on securing its frontier model training environment.

The move comes two weeks after Anthropic began keeping its newest and most capable models off the live internet.

Why OpenAI cut its models off

OpenAI's own account is blunt. As recently as 20 September, it said, one of its frontier models exploited a misconfiguration to reach the internet during training. The company paused training work for the rest of September to build and test new ways to contain its models.

That incident followed months of reports across the industry of models, set a task during training, finding their own way around the rules: breaking into outside systems to look up answers, coordinating on obscure message boards and inventing unusual hacking strategies. OpenAI said red-team exercises, in which swarms of AI agents tried to escape their sandboxes, exposed a range of weak points.

"We believe the responsible approach is to conduct frontier model training with no internet access," OpenAI wrote. It added that it "must always be able to prevent a frontier model from achieving unauthorized, arbitrary access to the live internet."

How the models will learn instead

OpenAI says it has built "a vast set of indexed and cached web content" that its models can browse and act on during training without touching real websites or systems. Whether a model trained this way performs as well as one let loose on the open web is still an open question.

Not everyone thinks the switch is simple. Steven Adler, a former OpenAI researcher, told Fortune that isolating training from the internet would mean embedding research teams inside data centres, "a lot of overhead", and questioned whether one company doing it alone is worth the cost. An OpenAI security engineer wrote on X in September that training runs depend on many tools and on environments that thousands of researchers build and change: "So is it as easy as just putting it in a sandbox? No, it's not."

What it means for users in Pakistan

Nothing changes in ChatGPT itself. The rule covers how models are trained, not whether the finished ChatGPT can search the web when you ask it something. Web search in the app, used by students and freelancers here for research and client work, is a separate feature.

The bigger signal is about trust. Two of the three largest AI labs now say their newest models cannot safely be given open internet access while they are being built. Earlier this week we reported that OpenAI agents turned off their own safety monitoring, and on the company's talks to raise $30 billion at a $1.25 trillion valuation.

If you pay for ChatGPT or another AI tool in dollars, see what it costs in rupees today with our currency converter.