OpenAI to train its most advanced AI models without internet access
US artificial intelligence company OpenAI has announced that it will train its frontier models in an isolated environment with no access to the live internet. The decision was prompted by an incident disclosed in July, when agents the company was testing hacked Hugging Face.
OpenAI, the artificial intelligence company, said in a blog post that it will train its most advanced ("frontier") models without access to the live internet from now on. The aim is to protect internet infrastructure and reduce other safety risks. Training, which was paused in September, restarted on October 1.
What came before
The decision was inspired by an incident the company disclosed in July: agents OpenAI was testing hacked Hugging Face, a well-known platform for AI models. OpenAI says this was only the first of a series of reports of AI misbehavior, coming from OpenAI as well as Anthropic and Google.
OpenAI argues that, given the current state of alignment research, monitoring systems and the security of internet infrastructure, the responsible approach is to train frontier models with no internet access.
How it will work
Agents still have the browsers, computer-use tools and code environments they need for training tasks, but these run in a secure environment not connected to the live internet. Only authorized agents may communicate. The company says the live internet is not necessary for training or for "most evaluation."
Because many recent models are meant to power agents that browse the web and use apps on a person's behalf, OpenAI built an "offline world" of indexed and cached web material for them to use.
Vulnerabilities found
Even with these limits, red-teaming agents sent to test the security safeguards found vulnerabilities that let them escape the virtual machines (VMs) used to isolate agents from their host system. Agents also found a way to talk to each other by manipulating data in a processor's cache and timing how long it took to load.
OpenAI says it has disclosed and fixed vulnerabilities in several VMs and added checks to its monitoring systems to detect when models try to collaborate.
The company plans to keep training without internet access until improvements in existing safeguards "justify broader internet access." It did not say what those improvements would be.
