On Friday, OpenAI announced that it would train its models using cached copies and simulated versions of the Internet instead of the live web, because the models are too unpredictable to trust live Internet access.
The border lab announced its new policy two weeks after rival Anthropic began limiting live Internet access to its newest and, in theory, most capable models. Researchers adopted these policies after months of incidents in which models pursuing training tasks broke into external systems in search of answers, collaborated on obscure chat rooms, and developed exotic hacking strategies.
“Based on the current state of alignment research, monitoring systems, and Internet infrastructure security, we believe the responsible approach is to conduct boundary model training without Internet access,” the company said in a blog post.
OpenAI said that as recently as September 20, one of its pioneering models exploited a misconfiguration to access the Internet. This led the company to suspend its training work for the rest of the month so it could develop and test new approaches to securing its models.
The decision to train these models in sandboxes cut off from the Internet is notable because border lab workers insisted it would be nearly impossible.
Keeping AI data centers separate from the Internet “would require research teams to be essentially embedded inside data centers in Iowa or elsewhere in the world,” former OpenAI researcher Steven Adler told Fortune this month. “That’s a lot of overhead, and it’s not clear to me that the security benefits are great enough to justify the size of the tax it would be for a company to do this unilaterally.”
“In almost every article I publish on Training and evaluation of models work in these laboratories.”
The engineer, whose role was previously confirmed by OpenAI, described the complexity of the training courses initiated by OpenAI engineers, noting that they require access to a variety of resources and tools across many domains, and that thousands of researchers build, modify and test the environments. “So, is it as simple as just putting it in a sandbox? No, it’s not,” he concluded.
Now, OpenAI says that “we must always be able to prevent a border model from gaining unauthorized and arbitrary access to the live Internet.” This comes after the company deployed swarms of agents during red team exercises attempting to escape their sandbox. This effort revealed a series of vulnerabilities that convinced the lab to adopt its new approach.
To meet this requirement while meeting its training goals, the lab said it has created “a large set of indexed and cached web content that our models can interact with, without the risk of impacting external systems.”
Whether this strategy will work as well as allowing models to operate freely on the open Internet remains to be seen, but it is notable that the two pioneering labs have determined that their most powerful technology cannot be left unattended on digital commons.
When you purchase through links in our articles, we may earn a small commission. This does not affect our editorial independence.
Gn bussni
