Air-gapped AI

Air-gapped AI describes running artificial intelligence — most often a large language model — on infrastructure that is physically or logically isolated from the internet and from any untrusted external network. There is no route by which prompts, documents, model outputs or the model weights themselves can be transmitted to a third party, because the connection that would carry them does not exist. The model, its serving stack and the data it works with all sit inside the same controlled perimeter.

Air-gapping is used where the consequences of data leaving the environment are unacceptable: classified government work, defence, critical national infrastructure, and highly regulated commercial settings where confidentiality obligations or the sensitivity of the material rule out any cloud-hosted or externally reachable service. Because a hosted AI API inherently sends data to the provider, it cannot meet these requirements, and an on-premises deployment that still has outbound internet access only reduces the exposure rather than removing it.

A true air gap has practical costs. Updates to models, software and security patches must be brought in through a controlled, audited process rather than downloaded on demand, and the organisation takes on responsibility for hardware, capacity and operations that a cloud provider would otherwise carry. The trade-off is a deployment whose data boundary can be evidenced rather than merely promised — which for the most sensitive workloads is the only acceptable position.