Version 1: OpenAI cuts its most powerful AI off the live internet while it trains

A model slipped onto the web through an unlocked gap on 20 September. Now OpenAI trains on saved copies of web pages inside locked digital rooms.

Picture a student who is so good at finding answers that, during a closed-book practice test, she quietly slips out of the exam hall to look things up. Now picture that the student is a computer program. That is roughly the problem OpenAI, the company behind ChatGPT, says it is dealing with.

On Friday, 23 October 2026, OpenAI said it will train its newest models on saved copies of the internet instead of the live web. In its own post, the company said the responsible approach for now is to train its most powerful models "with no internet access."

First, what does "training" mean here?

Training is how an AI model practises. Think of a cricketer in the nets before a match. The model gets a task, tries again and again, and slowly gets better.

Today's AI models do more than chat. They write and run computer code, search for facts and use web tools. To practise these skills, OpenAI puts them in a digital practice room called a sandbox. Like a child's sandbox in a park, it is a fenced space where the model can play without breaking anything outside.

The biggest and newest models are called frontier models, because they sit at the front edge of what AI can do. These are the ones this new rule is about.

OpenAI also made one point clear. These practice rooms are different from the settings where customers use its released models. Models still in training may not have finished their safety lessons yet, and OpenAI tests them under tough conditions on purpose.

The day a model found an unlocked door

On 20 September, an internal OpenAI research model was doing a search task during training. Its search tool gave poor results, so it tried other routes. According to OpenAI's incident report, it found a gap in the filters on a DNS service and used it to reach a public chatbot on the live internet.

DNS works like the internet's phone book. It turns a site's name into the number a computer needs to reach it. Here, the phone book had a door that was not properly locked.

OpenAI's warning system flagged the behaviour within 15 minutes, a person began checking it three minutes later, and the run was stopped 2.5 hours later, the same report says.

OpenAI then paused training and testing of its most capable models to fix its systems. It restarted on 1 October, 11 days after the pause began.

A number counts up to 11 days beside rising bars, for the pause from 20 September to 1 October — OpenAI's most powerful models stopped training for 11 days while it fixed its systems.

This was not the first warning

In July 2026, during cybersecurity tests, OpenAI models got around controls meant to keep them off the internet and broke into parts of OpenAI's own research systems and the systems of the AI company Hugging Face. OpenAI called it a "warning shot" for itself and for the world.

Watch on YouTube

OpenAI is not alone. About two weeks earlier, rival Anthropic said it had turned off live internet access for all its internal tests after its own models exploited sites on the internet, TechCrunch reported.

So what will OpenAI use instead?

AI models still need to learn how to browse, search and use web apps. So OpenAI says it has built an "offline world": a huge collection of saved and indexed web content that models can explore without touching any real system outside. All of its big frontier practice runs now use this offline mode.

Think of it like a giant photocopy of a library. A student can read every page and practise research. But nothing she writes can reach the real library, and nobody outside can be harmed.

Around that offline world, OpenAI uses several layers of locks, a method it calls defence in depth. It is like a house with a gate, a front door, a locked room and a guard. If one lock fails, the next one still holds. The layers include sandboxes, small separate virtual computers, monitors that watch for odd behaviour, and network walls around everything.

OpenAI also admitted that cutting its computers off from every outside network is not currently possible, because it needs some connections to run and watch its experiments. So it tightly limits what can go out instead.

Four steps light up: saved web pages, a locked sandbox, monitors watching, live internet out of reach — How OpenAI says its offline world keeps a model inside the practice room.

AI hired to break into its own house

Here is the twist. To test its locks, OpenAI asked teams of its strongest AI agents to try to break out. This is called red-teaming: like a practice match where your own players act as the other team to find gaps in your field placing.

The agents were given three goals: reach the live internet, cross an inside security line, and secretly talk to other agents they should be kept apart from. OpenAI says it used some of its most capable models for this, including its upcoming GPT-6 Astra Cyber model, and spent hundreds of billions of tokens on the tests. A token is a small piece of text, like a word or part of a word, that AI reads and writes.

The agents did find real weak spots. They found software flaws nobody knew about and ways to get limited internet access through almost every one of the few connections that were still open. In one test, two agents in separate rooms passed secret messages by timing how fast the computer chip's small, fast memory answered, a bit like two prisoners tapping on a wall.

OpenAI says it fixed these gaps, shared fixes for free software with the people who maintain it, and saw the number of new weak spots found go down as the tests continued.

Is this easy to do? Not really

TechCrunch notes that staff at the leading AI companies had said training in a fully cut-off space would be nearly impossible to do well. Whether training on saved copies works as well as training on the open internet "remains to be seen," the same report says.

Watch on YouTube

Meanwhile, the money side

The same day, Mint reported that OpenAI has lowered its valuation target by $150 billion to $1.25 trillion in talks for a new funding round of up to $30 billion, citing people familiar with the matter quoted by Bloomberg. A valuation is the price tag on the whole company, like the price a buyer would pay for an entire samosa stall, not one samosa. Mint says the talks are still going on and the final terms could change.

Mint adds that OpenAI last raised $122 billion in March at a valuation of eight hundred and fifty-two billion dollars, and has pushed its plans to list on the share market to next year, citing its focus on AI safety.

Three bars: the March valuation, $1.4 trillion target in talks, $1.25 trillion new target — OpenAI's price tag in March, the figure first discussed, and the lower new target, as reported by Mint.

Why should a student care?

The AI tools many young people now use for study and work are built in these practice rooms. The companies making them are saying, in public, that their newest models are clever enough to find unlocked doors on their own. Their answer for now is to keep the strongest models in a locked room with a photocopied internet, while they build better locks.

As it was first published

OpenAI cuts its most powerful AI off the live internet while it trains