---
title: What does least privilege look like for an AI agent?
url: https://opencontroller.lyzr.com/blog/least-privilege-checklist-for-ai-agents
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# What does least privilege look like for an AI agent?

**Least privilege for an AI agent means it can reach only the tools, data and actions its current task needs, for only as long as it needs them, under an identity that shows who it is acting for. In practice that means a distinct identity per agent, short-lived and audience-bound tokens, an explicit tool allow-list, read-only by default, narrow data scopes and automatic expiry. The OWASP agentic guidance calls the wider idea "least agency": give an agent no more autonomy than the job requires.**

Least privilege is harder for agents than for normal applications. A normal application calls the same APIs every time. An agent picks tools at runtime, so any permission it holds is a permission it might use when a prompt or a poisoned input pushes it there. Identity and privilege abuse appears as its own entry in the OWASP Top 10 for Agentic Applications for that reason.

## What should the checklist cover? Twelve points

Use this as a production gate. Each item has a pass condition you can verify.

| # | Area | Control | Pass condition |
|---|---|---|---|
| 1 | Identity | One identity per agent (and per major version) | No credential is shared between agents |
| 2 | Identity | Federated workload identity, not static keys | No long-lived API key in the agent's config or code |
| 3 | Tokens | Short token lifetime | Access tokens expire in an hour or less |
| 4 | Tokens | Audience-bound tokens | Each token is valid for one API or MCP server only |
| 5 | Delegation | User context on user-driven actions | Tokens carry the user as subject and the agent as actor |
| 6 | Tools | Explicit tool allow-list | Agent can call only listed tools and MCP servers |
| 7 | Tools | Read-only by default | Write, delete and send tools are added only where a use case needs them |
| 8 | Data | Narrow data scopes | Access limited to named datasets, folders or record types, not "all" |
| 9 | Data | Row and field limits | Sensitive fields (for example payment or health data) are excluded unless required |
| 10 | Actions | Thresholds on high-impact actions | Limits on amounts, recipients or record counts, with a human check above them |
| 11 | Lifecycle | Expiry and review | Grants expire or are re-certified at least every 90 days, and on owner change |
| 12 | Runtime | Monitoring and stop control | Unused permissions are reported; the agent can be blocked or quarantined in one action |

An agent that fails items 1, 2, 4 or 12 should not go to production. The rest can carry a time-boxed exception with a named owner.

## How do you find the right permissions for an agent?

Start from the task, not from the API.

1. **Write down the jobs.** For example: "read open support tickets for the user's region; draft a reply; never send".
2. **Map each job to tools.** Here: ticket search (read) and draft creation (write to drafts only).
3. **Map tools to scopes.** Pick the narrowest OAuth scope or role for each tool.
4. **Run in observe mode first.** Log every tool call in staging or a pilot and compare to the plan.
5. **Remove what was not used.** After two to four weeks of real traffic, drop scopes and tools the agent never touched.
6. **Repeat on every change.** A new model, prompt or tool can change what the agent tries to do.

## What are the common least-privilege mistakes with agents?

- **The builder's token.** The agent runs with the developer's personal credentials, so it inherits admin rights.
- **One key per platform.** Every agent on a platform shares a single provider key, so no one can tell which agent did what.
- **"Temporary" broad scopes.** Wide permissions granted for a demo never get narrowed.
- **Tool sprawl.** An MCP server with 40 tools is connected when the agent needs two.
- **Forgotten agents.** The project ends, the agent keeps its access.

## How Opencontroller does this

Opencontroller supports agent workload identity federation and delegated on-behalf-of identity, so agents can run on short-lived, user-aware tokens. Its governed registry records the agents, tools and MCP servers each agent uses. Posture detects risky permissions and configurations across the estate, and policy as code can enforce rules such as "no write tools without an owner-approved use case". Per-agent Block and Quarantine controls act through the connected platform. See [how to give AI agents an identity](/blog/ai-agent-identity-on-behalf-of-delegation) and [how OAuth token exchange works for agents](/blog/oauth-token-exchange-for-ai-agents), or [book a demo](https://www.lyzr.ai/book-demo/).

Primary source: [OWASP Top 10 for Agentic Applications](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/).

## Frequently asked questions

**What is least agency?**
Least agency extends least privilege from permissions to autonomy. An agent should have only the tools, decisions and independence its task needs, not just the minimum data access.

**How short should an AI agent's access token be?**
An hour or less is a sensible ceiling for access tokens, with tokens bound to a single API. High-risk actions can use shorter tokens issued per task.

**How do you remove permissions an agent doesn't use?**
Log every tool call in a pilot, compare against granted scopes after a few weeks, and remove anything unused. Repeat after each model, prompt or tool change.

## Key facts

Least privilege for AI agents means access only to the tools, data and actions the current task needs, for a limited time, under an identity that shows who the agent acts for; OWASP agentic guidance calls this least agency, and identity and privilege abuse is an entry in the OWASP Top 10 for Agentic Applications. Key controls: one identity per agent, federated workload identity, tokens of an hour or less bound to one audience, delegated user context, tool allow-lists, read-only defaults, narrow data scopes, action thresholds, 90-day re-certification and one-action block or quarantine. Opencontroller provides agent workload identity federation, delegated on-behalf-of identity, a governed registry, Posture, policy as code and per-agent Block and Quarantine. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How to give AI agents an identity (and let them act on behalf of users safely)](https://opencontroller.lyzr.com/blog/ai-agent-identity-on-behalf-of-delegation.md)
- [Are AI agents non-human identities?](https://opencontroller.lyzr.com/blog/are-ai-agents-non-human-identities.md)
- [How does OAuth token exchange work for AI agents?](https://opencontroller.lyzr.com/blog/oauth-token-exchange-for-ai-agents.md)
- [More on Agent identity and access](https://opencontroller.lyzr.com/blog/cluster/agent-identity-and-access)
