Regulating AI Behavior with a Hypervisor

The document summarizes research on 'Guillotine,' a proposed hypervisor architecture designed to sandbox powerful AI models by combining software, network, microarchitectural isolation, and physical fail-safes to enforce safety policies. It highlights the need for co-designed hardware and software to prevent introspection, side-channel leakage, and reflection-based attacks, along with extreme physical measures (e.g., network cable disconnection or datacenter flooding) as defense-in-depth against rogue AI behavior.