A host command is a command that executes on the local operating system, not inside a container, sandbox, or other constrained workspace. In security terms, the host is the real machine that owns files, processes, network settings, and credentials. If a tool is only supposed to work in isolation, a host command means that isolation has been crossed.
This matters because host execution turns a limited tool into a direct system actor. In attacks, sandbox escapes and command-injection flaws may let untrusted input trigger host commands, which can expose secrets, alter files, or launch additional tools. Defenders try to prevent this with strong sandboxing, OS-level isolation, least privilege, and permission prompts for risky actions. If an AI agent or automation platform can run host commands, it should be treated as potentially dangerous until the boundary is proven reliable.