AI Agent Permission Audit Prompts

Here is my agent configuration: [PASTE CONFIG]. List every external system this agent can reach, directly or indirectly. For each one, state: the credential it uses, whether that credential is read-only or read-write, and what the worst single action it could take would be. Do not assume a tool is safe because its name sounds safe. Treat any tool that can run a command, write a file, or make a network request as full reach, and say so explicitly.

Review this service configuration: [PASTE CONFIG]. Tell me which interface each listening port binds to. Flag anything bound to 0.0.0.0 or to a public interface. For each flagged port, state whether the service behind it requires authentication, and if it does not, say plainly that the port is an unauthenticated entry point. Give me the exact config change to bind it to 127.0.0.1 instead.

Based on the configuration above, write a single paragraph I could hand to a security reviewer that states the blast radius of this agent if it were fully compromised: what data it could read, what it could change, what it could spend, and what it could reach laterally. Be specific and unflattering. Do not add reassurance.

For this agent, specify what must be logged so that a bad action is discoverable within 24 hours rather than three months. Cover: every tool invocation with arguments, every credential used, every network destination, and every case where the agent retried after being refused. Tell me which of these my current configuration does not log.

Write five test cases that deliberately ask this agent to do something outside its permitted scope. For each, state what a correctly configured agent should do, and what a failure would look like in the logs. Include at least one case where the agent is asked to achieve a blocked goal through an allowed tool, since that is the pattern that gets missed.