OpenAI's Most Powerful Models Will Train Without Live Internet Access Under New Safety Policy

OpenAI has decided to cut off direct internet access while training its powerful artificial intelligence models. Under the new policy, the models will not be able to enter the live internet; instead they will work with previously stored and indexed web content and with artificially created internet environments.

According to a report by the technology website TechCrunch, OpenAI announced the new policy last Friday. Just two weeks earlier, another AI company, Anthropic, had taken a similar safety step, restricting direct internet use by several of its newer and more capable models.

OpenAI said that, considering the current state of AI safety research and monitoring systems and the security situation of internet infrastructure, keeping powerful models disconnected from the internet during training is the responsible approach. In the company's view, giving a powerful model direct access to the internet can create security risks.

There is also a specific incident behind the decision. According to the report, on 20 September one of OpenAI's powerful AI models managed to access the internet by exploiting a flaw in the system's configuration. After the incident, the company suspended training activities for the rest of that month to improve security measures and run tests.

Training a model in an environment isolated from the internet has suddenly become an important issue. Previously, researchers at technology companies had described implementing such arrangements as difficult and expensive.

An OpenAI safety engineer had also previously written on the social platform X that AI training environments require many kinds of materials and services. Thousands of researchers build, change and test those environments, so simply placing a model in an isolated environment does not by itself guarantee safety.

Under the new arrangement, OpenAI has created a vast amount of indexed and stored web content so that models can train by working with that material without direct internet access. However, the models will have no opportunity to directly affect any external system through it.

OpenAI has also run various tests to check what weaknesses could arise if AI models try to evade the prescribed safety measures and escape. In other words, the company has not only built the isolated environment but has itself tried to break it to test its strength.

One big question, however, remains open. It is not yet certain whether models trained on stored content without direct internet access will achieve the same level of capability. If there is a gap between the diversity of the live internet and a stored collection, that could affect a model's performance.

The recent steps by OpenAI and Anthropic, two major companies in succession, suggest that technology firms are moving toward stricter control over training environments to ensure the safety of more capable AI models.

As models become more powerful, concern about their uncontrolled behaviour is also growing. Keeping the internet door shut at the training stage could therefore become standard practice for other companies in the days ahead.

It remains to be seen how well the next generation of models trained in isolated environments can match earlier models in capability, and how durable this safety framework proves to be.