You Cannot Trust Your Agent (and How to Isolate It)
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
A developer describes running AI code agents inside a kernel-level sandbox built with bwrap (bubblewrap) on Linux, so the agent sees the root filesystem read-only, only the target project with read-write access, and home-directory config writes confined to temporary overlays.