Give AI Agents Only the Permissions They Need

Environment and inputs: A test agent identity, its effective tool permissions and a disposable target with one allowed read and one forbidden write.

An agent with broad access can turn an incorrect answer into a damaging action. OWASP describes excessive agency as a combination of too much functionality, permission, or autonomy. Control the application’s actual capabilities rather than assuming that a careful prompt will prevent every misuse.

What matters

• Give read-only tools to tasks that only need to inspect information.
• Use narrowly scoped service identities instead of shared administrator credentials.
• Require meaningful confirmation for consequential actions and keep auditable action records.

Practical example: A mailbox assistant can draft a reply but cannot independently send messages to arbitrary external recipients.

One action today: Remove one unnecessary tool or privilege from an existing agent integration.