It is 2026, and it is common to run AI agents in a terminal or an integrated development environment for coding tasks.

The problem is, AI agents can be insecure to run directly on your machine: only in Summer 2026 were several incidents reported.¹ I have recently experienced a strange situation myself when Antigravity, an AI agent from Google, has asked me if it was allowed to run tests from the Tests/ directory one level above the current working directory, because it has looked there for unknown reason. Of course, many software packages look for different things on the file system, for example, for a C compiler, etc. However, they usually check some predefined system locations like /usr/bin on Linux.