Definition
Excessive agency is the security condition created when an AI system has more functionality, permission, autonomy, or action scope than its job requires. The term is used by OWASP for a major class of risks in applications built around language models.
Excess can appear in several places. An agent may have tools it never needs, credentials with broad access, permission to affect every customer instead of one, or authority to execute an irreversible action without review. The model does not need to be malicious for that authority to become dangerous; a hallucination, prompt injection, or ordinary software bug can be enough.
Reduce the authority itself: expose fewer tools, grant narrower and shorter-lived credentials, constrain the resources each action can touch, separate proposing from executing, and place human approval before high-impact or irreversible effects.
Distinguish it from nearby terms
Agency is the capacity to pursue a goal and take action. Excessive agency is avoidable overreach in that capacity. Least privilege is the design principle used to reduce it, while a guardrail is one possible control and may or may not enforce a real boundary.
Check your understanding
A calendar assistant only needs to draft an invitation for review, but it can delete every event in the company. That unused permission is excessive agency even before anything goes wrong.