What is Airgapping?

What it is

An airgapped system has no connection to the internet or to any other network outside its own walls. The name comes from the idea of a literal gap of air: no cable, no wireless link, nothing for data to cross unless a person carries it across.

The technique is old. Governments, banks, power plants and labs have long used it for their most sensitive systems.

Why it comes up with AI

Most useful AI agents need the outside world. They browse, call APIs, read email and run code. Every one of those connections is a path out. If a model misbehaves, whether from a bug, a mistake or an attack such as prompt injection, a connected model can act on other systems. An airgapped model can only act on what is inside the gap.

That is why some security researchers argue that models with strong hacking ability should be tested and run in airgapped setups by default, and connected only when the risk is understood.

How it differs from a sandbox

An agent sandbox is an isolated environment with limits on what the agent can reach. A sandbox can still allow some network access, such as one approved website. An airgap is the strictest version: no network path at all. A sandbox is a fence with a gate. An airgap is a moat with no bridge.

The tradeoffs

• Safety. It removes whole classes of risk, because the model cannot send data out or attack something it cannot see.
• Usefulness. An airgapped model cannot check live information, use online tools or take actions for you, so many products would not work.
• Not perfect. Attackers have moved data across gaps using removable drives and even signals from hardware. People carrying files in and out are the usual weak point. Containment also depends on the sandbox or setup being built correctly.

What to take from it

For everyday use, a full airgap is rarely practical. The lesson to copy is the principle behind it: give an agent the least network access its task needs, and log what it does.