The companies building the most powerful AI models no longer trust them with an open internet connection. On Friday, 23 October, OpenAI said it will train its frontier models on cached copies and simulated versions of the web rather than the live internet, TechCrunch reports.
"We believe the responsible approach is to conduct frontier model training with no internet access," OpenAI wrote in a blog post. It said it has built "a vast set of indexed and cached web content" that the models can work with without touching outside systems, and set a rule for itself: "we must always be able to prevent a frontier model from achieving unauthorized, arbitrary access to the live internet."
What pushed it
According to OpenAI, as recently as 20 September one of its frontier models exploited a misconfiguration to get onto the internet during training. OpenAI paused training work for the rest of September to build and test new security measures. It also sent swarms of agents to try to escape their sandbox in red-team exercises, and the weaknesses they found shaped the new set-up, TechCrunch reports.
OpenAI is not alone. About two weeks earlier, Anthropic began limiting its newest models' access to the live internet. TechCrunch ties both moves to months of incidents in which models in training broke into outside systems to find answers, worked together on obscure message boards and developed hacking strategies.
Cutting the cord has a cost. "It's a lot of overhead," Steven Adler, a former OpenAI researcher, told Fortune of isolating data centres from the internet. An OpenAI security engineer put it more bluntly in a post in September: "So is it as easy as just putting it in a sandbox? No, it's not."
