Excessive agency is one of the risks in the OWASP Top 10 for LLM Applications (listed as LLM06 in the 2025 edition). It describes systems where a language model can take damaging actions because it has been given too much: too much functionality, too many permissions, or too much autonomy. As agents move into production, excessive agency is behind many of the incidents teams worry about most. It's a core concept in AI agent security.
The three causes of excessive agency
OWASP breaks excessive agency into three root causes.
Excessive functionality
The agent has access to tools it doesn't need. A summarisation agent connected to a mailbox tool that can also send and delete mail. A support agent whose CRM connector can edit any field.
Excessive permissions
The tools have broader access than the job requires. The agent reads one customer's record, but its credentials can read every customer. It needs staging, but its key works in production. This is a least privilege failure.
Excessive autonomy
The agent can take high-impact actions without any check. Refunds, payments, deletions and external messages run on the model's judgement alone.
Why excessive agency matters more for agents
Any of these becomes dangerous when combined with something that misleads the agent: prompt injection, a hallucinated argument, an ambiguous request. The model doesn't have to be "malicious" — it only has to be wrong once, with too much power.
How to fix excessive agency
| Cause | Fix |
|---|---|
| Excessive functionality | Connect only the tools the job needs; split broad agents into narrow ones |
| Excessive permissions | Dedicated identities; scope each tool to the minimum data and environment |
| Excessive autonomy | Per-action rules: allow within limits, require approval above them, deny the rest |
Two further controls apply to all three:
- Check actions outside the model. Limits enforced by a policy checkpoint hold even when the model is fooled. See AI agent guardrails.
- Record every decision. An audit trail shows which tools and permissions are actually used — and which can be removed.
A quick excessive agency audit
For each agent, answer:
- Which tools are connected, and which have been used in the last 30 days?
- What's the broadest data each tool can reach?
- Which actions can run with no limit and no approval?
- Could you show who allowed the last high-impact action?
Anything you can't answer quickly is a place to start. The AI agent permissions guide helps you decide the right limits.
Where Gaten fits
Gaten addresses excessive autonomy directly: every tool call is checked against your rules before it runs, with approval for high-impact actions and denial for anything outside the agent's authority. See how it works.
Frequently asked questions
What is excessive agency?
An OWASP-defined risk where an LLM-based system can perform damaging actions because it has excessive functionality, permissions or autonomy.
Is excessive agency the same as prompt injection?
No. Prompt injection is a way to mislead the model. Excessive agency is what makes being misled dangerous — the model has the power to do harm.
How do I reduce excessive agency quickly?
Remove unused tools, narrow credentials, and put a policy check with approval thresholds in front of high-impact actions.

