Python Tool Lets Developers Audit Exactly Which Files an AI Coding Agent Touches

A developer has published a lightweight, dependency-free Python script called fence_check.py that audits the filesystem activity of AI coding agents during task execution. The tool addresses a subtle but common risk: agents successfully completing tasks while quietly modifying files outside their intended workspace, such as editing parent directories or writing to unintended locations. Rather than relying on privileged system tools like strace or eBPF, the script snapshots directory trees before and after an agent run and diffs the results against a defined allowlist. Any file change detected outside permitted paths is flagged as a breach, giving developers a repeatable, low-cost way to verify that their assumed sandbox boundaries actually hold. The author emphasizes that the threat model is not a malicious AI but an ordinarily helpful one operating with overly permissive tool access.