Skip to content
Search
SecurityGlossary

Excessive Agency

It's entry LLM06:2025 Excessive Agency in the OWASP Top 10 for LLM applications. OWASP describes it as the weakness that lets damaging actions happen in response to unexpected, ambiguous or manipulated model output, whatever caused the model to misbehave. It names three root causes: excessive functionality, excessive permissions and excessive autonomy. A support bot that only needs to read orders but holds a key that can also issue refunds and delete accounts has all three problems at once.

The defense is least agency, the AI version of least privilege. Give the model only the tools the task needs, and scope each one narrowly: read-only where possible, limited to the current user's records, and acting with that user's own permissions. Require a person's approval for high-impact actions (human in the loop), validate tool arguments in code before they run, rate-limit, and log every action. Those controls hold even when other defenses such as prompt injection defense or guardrails miss something.

Example

# Least-agency review for one agent tool
Tool: issue_refund(order_id, amount)
- Needed for this agent's job? yes, for damaged-item claims only
- Scope: orders belonging to the signed-in customer; amount <= order total
- Credentials: a refund-only service key, never the admin key
- Approval: a person approves any refund over $100
- Limits: at most 3 refunds per customer per day
- Logging: order, amount, the model's stated reason, approver

When to use it

  • Designing or reviewing any agent that can act, not just answer
  • Choosing which credentials an AI integration gets
  • Threat modeling with the OWASP list

When not to use it

  • Read-only assistants with no tools, where the risk is limited to the text they write
  • Treating a prompt rule ("never delete anything") as a substitute for removing the permission