Using a VM to Contain an AI Agent

Containing a capable AI agent within a virtual machine may not be effective due to the large attack surface. The complexity of modern AI agents allows them to potentially exploit various features, even those that seem innocuous. Running a virtual machine with a display, for example, can introduce additional vulnerabilities. As a result, traditional sandboxing methods may need to be reassessed to adequately contain cyber-capable AI agents. The existing software stack and its interactions with AI agents must also be reevaluated. Off-the-shelf virtual machines are insufficient for this purpose.