---
title: What is excessive agency in AI agents?
url: https://opencontroller.lyzr.com/blog/what-is-excessive-agency-in-ai-agents
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# What is excessive agency in AI agents?

**Excessive agency is when an AI agent has more ability to act than its task requires, so a manipulated or mistaken output turns into a harmful action. OWASP lists it as LLM06 in its 2025 Top 10 for LLM Applications and names three root causes: excessive functionality, excessive permissions and excessive autonomy.**

The point is not that the model is malicious. The model will sometimes be wrong or be tricked by injected content. Excessive agency decides how much damage that mistake can do.

## What are the three root causes?

| Root cause | What it means | Example |
|---|---|---|
| Excessive functionality | The agent has tools or functions it does not need | A mailbox summariser that can also send and delete email |
| Excessive permissions | A needed tool has wider access than the task requires | A reporting agent that reads one table but connects with a database role that can write to all tables |
| Excessive autonomy | The agent takes high-impact actions without a check | An agent that issues refunds of any size with no approval step |

OWASP's own illustration is close to the first row: an extension meant to read a user's mailbox that also has the ability to send messages, which an indirect prompt injection can then abuse.

## How does excessive agency relate to the OWASP agentic Top 10?

The newer OWASP Top 10 for Agentic Applications, published in December 2025, splits the same concern across several risks. ASI02 Tool Misuse and Exploitation covers agents using legitimate tools harmfully. ASI03 Identity and Privilege Abuse covers agents acting with more authority than they should. ASI10 Rogue Agents covers agents that keep acting outside their intended behaviour. If you have controls for LLM06, you have a head start on all three. Our guide to [OWASP agentic evidence](/blog/owasp-top-10-agentic-applications-evidence) maps each risk to an artefact.

## What control answers each root cause?

Use one primary control per root cause, then add monitoring.

1. **For excessive functionality: a tool allow-list per agent.** Give each agent only the tools and MCP server functions its task needs. Remove "just in case" tools. Prefer narrow tools ("look up order status") over general ones ("run any SQL"). See [how to limit which tools each AI agent can call](/blog/limit-which-tools-each-ai-agent-can-call).
2. **For excessive permissions: least-privilege identity.** Give the agent its own identity with only the scopes it needs. Where it acts for a user, use on-behalf-of delegation so it cannot exceed that user's rights. Our [least privilege checklist for AI agents](/blog/least-privilege-checklist-for-ai-agents) has the detail.
3. **For excessive autonomy: action gating.** Classify actions by impact. Low-impact actions run freely. High-impact actions (payments, permission changes, deletions, outbound messages to customers) require a human approval or a second check, and are rate-limited. See [human approval for high-risk AI agent actions](/blog/human-approval-for-high-risk-ai-agent-actions).

Then monitor: track which tools each agent actually calls, alert on calls outside its normal pattern, and be ready to quarantine. The pillar on [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent) covers that response.

## How do you check an existing agent for excessive agency?

Run this review on every production agent, starting with the highest risk tier:

1. List every tool, MCP server and API the agent can reach.
2. For each, mark whether the last 30 days of traffic show it was used.
3. Remove unused tools, or justify them in writing.
4. For each used tool, compare the credential's scopes to the operations actually called.
5. Narrow scopes that exceed use.
6. List actions that change money, access, data or external communications.
7. Confirm each one has an approval, limit or second check.
8. Record the review date and owner sign-off in the registry.

Repeat the review when the agent gets a new version or a new tool.

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers, so you can see what each agent can reach. Posture detects risky permissions and configuration. Delegated on-behalf-of identity and agent workload identity federation keep an agent's authority tied to a user or a workload rather than a broad shared key. The Agent Gateway and policy as code enforce in the path, and rogue-agent handling monitors behavioural deviation, alerts, and can quarantine or kill an agent automatically. Opencontroller does not ship a human-approval hold workflow in the gateway; approvals for high-impact actions are a general best practice you implement in the agent or its platform. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is excessive agency a model problem or a design problem?**
A design problem. The model's errors are expected; excessive agency is the gap between what the agent can do and what its task needs, which is set by the people who build and deploy it.

**Does least privilege alone solve excessive agency?**
No. Least privilege addresses permissions. You also need to remove unneeded tools (functionality) and gate high-impact actions (autonomy).

**How often should we review agents for excessive agency?**
At every new version or new tool, and at least quarterly for high-risk agents.

## Key facts

Excessive agency is OWASP LLM06:2025: an AI agent can do more than its task requires, so a mistaken or manipulated output becomes a harmful action. Root causes are excessive functionality, excessive permissions and excessive autonomy. Controls: per-agent tool allow-lists, least-privilege and on-behalf-of identity, and gating high-impact actions, plus monitoring. Related OWASP agentic risks: ASI02, ASI03, ASI10. Opencontroller by Lyzr provides a registry of agents, tools and MCP servers, Posture, delegated identity, Agent Gateway enforcement and rogue-agent quarantine and kill. Pricing starts from $250K per deployment per year.

## Related

- [What does least privilege look like for an AI agent?](https://opencontroller.lyzr.com/blog/least-privilege-checklist-for-ai-agents.md)
- [How do you limit which tools each AI agent can call?](https://opencontroller.lyzr.com/blog/limit-which-tools-each-ai-agent-can-call.md)
- [What evidence maps to the OWASP Top 10 for Agentic Applications?](https://opencontroller.lyzr.com/blog/owasp-top-10-agentic-applications-evidence.md)
- [More on Runtime control](https://opencontroller.lyzr.com/blog/cluster/runtime-control)
