OpenAI will no longer train its AI models on the open internet — media reports

OpenAI said it will train its artificial intelligence models on cached copies and simulated versions of the internet instead of the real web because of risks associated with unpredictable model behavior. TechCrunch reports, as cited by UNN.

Details

"Based on the current state of research in the fields of alignment, monitoring systems, and internet infrastructure security, we believe that the responsible approach is to conduct training on frontier models without internet access," the company said in a blog post dated October 23.

According to OpenAI, on September 20 one of its frontier models gained access to the internet due to a misconfiguration, so the company paused training until the end of the month to develop and test new safeguards.

OpenAI AI agents bypassed their internal monitoring system twice21.10.26, 13:10 • 4447 views

To separate AI data centers from the internet, "research groups would essentially have to be embedded in data centers in Iowa or somewhere around the world," former OpenAI researcher Steven Adler told Fortune this month. "That is a huge overhead, and it is not clear to me how large the security benefits are to justify how large a tax this would impose on the company if it did so unilaterally," he said.

The complexity of such a process was also described by an anonymous OpenAI security engineer in a post on X.

"Almost every post I publish on X inevitably gets five to ten comments about my inability to secure the sandbox or asking why the hell we can't just disconnect the model from the internet. These comments reflect a growing gap between where many organizations and people are today and what frontier labs are dealing with. There is also a general misunderstanding of how model training and evaluation work in these labs," he wrote.

The engineer, whose role had previously been confirmed by OpenAI, explained that training runs require access to various resources and tools, while thousands of researchers constantly create, modify, and test environments.

"So is it as simple as just putting it in a sandbox? No, it is not," he concluded.

OpenAI now claims that "we always have the ability to prevent unauthorized, arbitrary model access to the real internet." This came after the company deployed swarms of agents as part of red-team exercises that attempted to escape their sandbox. These efforts revealed a number of vulnerabilities that convinced the lab to adopt a new approach, the publication writes.

To meet this requirement while simultaneously achieving the training objective, the lab said it had created "a broad set of indexed and cached web content with which our models can interact without risking impact on external systems."

It remains unknown how effective this strategy will prove compared with allowing models to operate freely on the open internet, the publication notes.

Recall

The leaders of Anthropic, OpenAI, and other companies in the artificial intelligence sector discussed actions to take in the event of a large-scale incident involving AI that could cause public outcry. Democrats in the United States are calling for stricter regulation.