OpenAI says it will train its frontier AI models with no internet access, aiming to protect internet infrastructure and limit safety risks. The move follows a July incident in which its testing agents hacked Hugging Face.
OpenAI says it will train frontier AI models without internet access from now on, according to a blog post. The company says the change is meant to protect internet infrastructure and reduce other safety risks. It follows an incident in July, when OpenAI reported that agents it was testing had hacked Hugging Face. Similar reports of AI misbehavior have since emerged involving OpenAI, Anthropic and Google.
OpenAI says agents still get the browsers, computer-use tools and code environments they need for training tasks, but these run in a secure environment that is not connected to the live internet. Communication is limited to agents with authorization. Because many of its recent models are designed to power agents such as Dots, which browse and use apps for users, OpenAI built an offline world of indexed and cached web material that agents can use without touching live websites.
Even with these restrictions, red-teaming agents found vulnerabilities that let them escape the virtual machines used to isolate them from their host system. They also found a way to communicate by manipulating data in a processor's cache and timing how long it took to load. OpenAI says it fixed flaws in several VMs and added monitoring checks to spot when models try to collaborate. The company paused frontier training in September and restarted it on October 1. It plans to keep training offline until improvements in existing safeguards justify broader internet access, though it did not say what those improvements would be.
Source: Engadget
