Is Your AI Agent Safe, or Just Configured to Appear Safe?

An AI agent that passed testing could gain write access to customer records after a single permission change, with its model and prompt left untouched. The article argues there is a critical difference between safety achieved by configuration and safety built in by default. Most agent platforms offer controls such as tool permissions, approval settings, and execution limits, but their effectiveness depends on developers and administrators configuring, maintaining, and enforcing them correctly. The author recommends starting from a restrictive baseline where high-impact actions stay unavailable until explicitly authorized, with unknown policy outcomes defaulting to denial. Instructions in a system prompt are not enough, so authorization should be checked by an external enforcement mechanism rather than left to the model's own reasoning.