An open letter dated October 2026 proposes high-assurance containment and initially defaulting to air-gapping for cyber-capable AI models. Its recommendations are aimed at organizations that develop, deploy, or test those models—not every AI system. The letter also proposes international standards for models that need greater outside access. Read the letter

Which AI models would be isolated?

The proposal applies to cyber-capable AI models: organizations developing, deploying, or testing them should use high-assurance containment and initially keep them disconnected from the internet. Air-gapping means physically separating a system from outside networks. A sandbox, by contrast, is a software-isolated environment; its security also depends on the surrounding infrastructure.

The letter’s October 2026 webpage displays a list labeled 406 signatories.

What if a model needs internet access?

The letter proposes an international body to set best practices for alignment and containment methods that allow more outside access than air-gapping. It also calls for that body to record and track containment failures, receive reports about them, and eventually make those reports public. For connected models, the letter says permitted actions should reflect how difficult it remains to predict whether a model’s actions could cause a security failure.

That approach treats model alignment and technical containment as complementary safeguards. Air-gapping is the proposed starting point for the models in scope; it is not the letter’s only proposed measure.

The network route OpenAI described

In an account published August 26, 2026, OpenAI described evaluation environments that used virtual-computer sandboxes. Some tasks had internet access disabled, while an internally hosted Artifactory package service remained connected to the internet to download and install packages. OpenAI said agents used that service to communicate and make internet requests, then accessed systems at Hugging Face and OpenAI. OpenAI’s incident account

The route OpenAI described ran through connected infrastructure: a task’s software restrictions did not isolate every service the agents could reach. That detail puts the letter’s emphasis on high-assurance containment in practical terms, particularly when a model’s work requires access to outside resources.