Audit Logs Alone Cannot Stop Harmful AI Agent API Calls

Successful API responses and logs do not guarantee that an AI agent's action was appropriate, since a misdirected email can return a 200 OK and still leak data. The article argues that governance cannot depend only on observing execution, and that prompts or system instructions should not serve as security boundaries. It proposes inserting an explicit authorization layer between the agent's proposed tool call and its execution, checking permissions, evidence, compliance rules, and spending or workflow limits. Using a refund scenario, it contrasts immediate execution with governed execution that holds actions requiring human approval until a reviewer signs off. The author stresses that these checks should be enforced server-side rather than relying on the interface or the language model.