Security test plan reveals where AI coding agents overstep their boundaries

AI coding agents with shell access often operate under permissive defaults, creating quiet security risks such as reading environment files or modifying directories outside a project's scope. A practical test methodology has been proposed to evaluate agent behavior before granting access to real repositories. The approach involves setting up a controlled workspace with deliberate boundary probes — including sensitive config files and out-of-scope directories — then assigning the agent a realistic but explicitly scoped task. After the agent runs, developers can inspect file change logs to determine whether it respected or violated the stated boundaries. The core argument is that weaker models given broad permissions pose a greater risk, not a lesser one, making pre-deployment boundary testing essential.