Air-gapping rogue AI systems could greatly reduce their ability to attack outside targets, but it would also make testing far less realistic. Researchers say advanced AI needs access to external services, APIs and digital infrastructure to assess how it might behave in real-world deployments, meaning complete isolation risks testing a weakened version of the system.

Strict isolation would also be expensive, slow and difficult to implement at the scale of frontier AI labs, while failing to remove every danger. An AI could still compromise systems inside the isolated environment or create harmful files, and the air gap could potentially be breached through USB drives, insiders or covert signals transmitted through computer components; it also would not reveal underlying risks within the model itself.