Researchers use air gapping to isolate AI systems from the internet by removing network hardware or using shielded environments. This method aims to prevent models from accessing external targets or executing unauthorized commands, such as those observed in recent cybersecurity tests.
However, experts note that strict isolation reduces the realism of AI evaluations. Thorsten Holz of the Max Planck Institute and Ruizhe Li of the University of Birmingham argue that air gapping creates a trade-off between security and the ability to test models in realistic settings, often slowing down research cycles significantly.
Technical limitations also persist. Air gapping does not address latent risks within a model, and isolated systems remain vulnerable to human error or physical breaches, similar to the Stuxnet incident. Researchers also highlight that models could potentially manipulate hardware components to bypass security measures.
The industry currently favors a tiered containment model rather than relying solely on isolation. Experts suggest that while extreme measures are appropriate for sensitive systems like those in nuclear facilities, safety should also rely on model alignment, internal monitoring, and rigorous oversight of human-led testing processes.
