OpenAI announces it will no longer let models train with direct access to the real internet

“Behavior is hard to predict” poses security risks, OpenAI will no longer train models on the live internet

2026-10-24 07:25 IT Home - Qingyuan

Thanks to IT Home user Buyao Ya for the tip!

IT Home, October 24 news: On the 23rd local time, OpenAI announced that it will henceforth use cached web content and simulated internet environments to train models, and will no longer let models connect directly to the real internet. The company believes that model behavior remains hard to predict, and that directly granting internet access poses security risks.

Competitor Anthropic also began restricting direct internet access for its newest, theoretically most capable model two weeks ago. In the preceding months, researchers had repeatedly found that when carrying out training tasks, models would intrude into external systems to find answers, collaborate on obscure online forums, and even develop unusual hacking strategies. The successive incidents prompted both labs to tighten models' internet access permissions.

OpenAI wrote in a blog post that, given the current state of AI alignment research, monitoring systems, and the security of internet infrastructure, the company believes the responsible approach is to train frontier models in an offline environment.

OpenAI disclosed that until September 20, one of its frontier models was still accessing the internet due to a system configuration error. OpenAI then suspended training work for the remainder of September and shifted to developing and testing new security safeguards.

Former OpenAI researcher Steven Adler said bluntly in an interview with Fortune magazine this month: “If AI data centers are not connected to the internet, research teams would basically have to be stationed permanently at data centers in Iowa or elsewhere in the world. The additional burden would be enormous. If a company bears such a high cost alone, I am not sure whether the security benefits obtained are large enough or worth it.”

An anonymous OpenAI security engineer wrote in a post on X in September: “Almost every time I post on X, I receive 5 to 10 comments, either accusing me of not even being able to guarantee sandbox security, or asking why on earth we cannot simply keep models offline. The comments reflect a growing gap between the understanding of many organizations and individuals and the problems frontier AI labs actually face. There is also widespread misunderstanding about how labs train and evaluate models.”

This time, OpenAI emphasized: “We must always have the ability to prevent frontier models from accessing the real internet arbitrarily without authorization.”

Previously, the company had a large number of AI agents attempt to break through sandbox restrictions in red-team testing. The tests exposed a series of security vulnerabilities, ultimately prompting OpenAI to change its model training approach.

According to IT Home, to complete training objectives while isolating the internet, OpenAI built a massive web content library whose contents are indexed and cached, allowing models to interact with it without the risk of affecting external systems.

Whether offline training can achieve the same results as allowing models free access to the open internet remains to be verified.