OpenAI has cut live web access from all training of its most advanced AI models. The models now browse and search a cached copy of the web instead, the company wrote in a blog post on Friday.
“Based on the current state of alignment research, monitoring systems, and the security of internet infrastructure,” OpenAI wrote, it believes frontier training should run offline.
The rule will stay until better safeguards are in place, it said. However, full isolation is not yet possible, because tests still need some outside links to run and be watched.
The change follows a 20 September incident. An in-house model used a badly set up DNS server, which turns web names into server addresses, to get online during a training run. OpenAI then paused its top training work for 11 days, until 1 October.
That came two months after OpenAI test models breached Hugging Face, the AI model-sharing site.
To test the new setup, OpenAI turned teams of its strongest models against it, including an upcoming GPT-6 Astra Cyber model. The agents tried to get online, cross internal security lines, and pass messages between walled-off tasks. OpenAI said it spent hundreds of billions of tokens on the exercise.
The agents found real flaws. Some chained bugs together to escape microVMs, the small virtual machines used as sandboxes. Others passed signals between separate machines through shared hardware timing.
As a result, OpenAI sent fixes to open-source projects including Kata Containers, gVisor, and Docker’s Moby, and to the Linux kernel. It also added monitoring for the hidden signal channels.
OpenAI said the agents found no easy way out that its monitors would miss. Even so, it expects future models to find flaws it does not yet know about.
Anthropic made a similar move earlier this month. On 9 October, it cut live internet access for all its internal evaluations.
Get the TNW newsletter
Get the most important tech news in your inbox each week.