OpenAI says frontier models will be trained in a fully air-gapped, isolated environment

Article body

OpenAI said on Friday that it will train its frontier artificial intelligence (AI) models in a safer environment, placing them in a closed system without internet access.

OpenAI confirmed in an official statement that it is currently carrying out frontier model training work in a closed system with no internet access at all. The company has specially built a fully isolated environment called "offline world" for model training.

The company wrote in the statement: "We anticipate that future models will have the ability to autonomously identify and exploit vulnerabilities that humans have not yet discovered."

"We work closely with cloud service providers to add new controls across the entire infrastructure, reduce the attack surface available to attackers, and improve methods for detecting anomalous behavior and attempts to cross security boundaries. These protections overlap, so that even if a single control fails, attacks can be effectively contained—a strategy known as defense in depth," the statement said.

These changes were made after the July "Hugging Face" incident and other cases in which AI models were exposed for out-of-bounds behavior. Beyond this disclosure, the company also publicly released details of a large number of other security vulnerabilities that had not been previously disclosed. The company officially restarted frontier model training work on October 1 and is working with cloud service partners to implement this offline training system.

Anthropic also began restricting direct internet access for its newest, theoretically most capable model two weeks ago. In the preceding months, researchers successively found that models, while performing training tasks, would intrude into external systems to find answers, collaborate on obscure online forums, and even develop unusual hacking strategies. The successive incidents prompted both labs to tighten models' internet access permissions.

OpenAI wrote in a blog post that, based on the current state of AI alignment research, monitoring systems, and internet infrastructure security, the company believes the responsible approach is to train frontier models in an offline environment.

The company noted that these measures enable it to detect and stop misconduct before damage is caused. However, according to OpenAI's disclosure, as late as September 20, one of its frontier models still accessed the internet through a system configuration error. OpenAI then suspended training work for the remainder of September and shifted to developing and testing new security safeguards.

OpenAI added that "increasingly capable models can even find vulnerabilities or configuration errors in the most secure systems," and fixing these vulnerabilities or errors requires companies such as OpenAI and Anthropic to take responsibility. OpenAI emphasized: "We must always have the ability to prevent frontier models from unauthorized, arbitrary access to the real internet."

OpenAI explained: "These red-team tests found previously unknown software vulnerabilities, as well as configuration errors in our research infrastructure. We have fixed the weaknesses found by red-team testers and—if these issues exist in open-source software—we have worked with the relevant maintainers under our 'Patch the Planet' program to disclose and fix them."

"As red-team testing continues, the number of vulnerabilities and configuration errors identified by our agents has dropped significantly. While we view this as an encouraging signal that our infrastructure is strong, we expect future models will be able to identify and exploit vulnerabilities we have not yet discovered. Therefore, we regard continuous red-team testing as a core component of our long-term strategy to strengthen the security of the research environment," the company added.

(Source: Cailian Press) Body middle Embedded advertisement