---
title: Are AI agents non-human identities?
url: https://opencontroller.lyzr.com/blog/are-ai-agents-non-human-identities
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# Are AI agents non-human identities?

**Yes, AI agents are non-human identities, but they are not just another service account. Like a service account, an agent authenticates with credentials rather than a person's login. Unlike one, it decides at runtime which tools to call, often acts for a specific user, and can change behaviour when its prompt, model or tools change. Govern agents with your non-human identity controls, then add delegation, intent and behaviour controls on top.**

Identity teams already manage thousands of non-human identities (NHIs): service accounts, API keys, OAuth apps, workload identities and certificates. Agents land in that same estate, which is why the question matters. The answer decides who owns agents, which tools see them and which policies apply.

## How do agents compare with other non-human identities?

| Attribute | Service account | Workload identity | AI agent identity |
|---|---|---|---|
| What it represents | An application or job | A specific running workload (pod, function, VM) | An autonomous component that plans and calls tools |
| Typical credential | Static key or password | Short-lived token from federation | Workload token plus delegated user tokens |
| Acts for a user? | Rarely; usually for itself | Rarely | Often, on behalf of the person who asked |
| Actions known in advance? | Yes, coded paths | Yes | No; chosen at runtime from the tools available |
| Can its behaviour change without a code deploy? | No | No | Yes, through prompt, model or tool changes |
| Usual owner | App team | Platform team | Often unclear: builder, business sponsor or platform |
| Main audit question | Which system called? | Which workload called? | Which agent, for which user, with what intent? |

The last three rows are why agents need extra controls. A service account's permissions describe what it will do. An agent's permissions describe only what it *could* do.

## Which existing NHI controls still apply?

Most of them. The OWASP Non-Human Identities Top 10 (2025) names risks that hit agents directly:

- **Improper offboarding:** agents left running after their project or owner is gone.
- **Secret leakage and long-lived secrets:** API keys in agent configs, notebooks and local MCP server files.
- **Overprivileged NHIs:** agents given broad admin scopes "to make the demo work".
- **NHI reuse:** several agents sharing one key, which destroys attribution.
- **Human use of NHIs:** people borrowing an agent's credentials to bypass their own limits.

Apply the same basics: inventory, owners, short-lived credentials, least privilege and regular access reviews.

## What extra controls do agents need?

1. **A distinct identity per agent and version.** Not one key per team. Attribution breaks without it.
2. **Delegated user context.** When an agent acts for a user, the token should carry both identities, for example through OAuth token exchange with an actor claim.
3. **Tool-level permissions.** Grant access to specific tools and MCP servers, not just APIs.
4. **Behavioural baselines.** Alert when an agent's tool calls, data volumes or destinations drift from normal.
5. **Change triggers.** Treat a new model, prompt or tool as a change that can require re-review, not just a code release.
6. **Runtime stop controls.** The ability to block or quarantine one agent quickly without revoking a shared credential that other systems use.

## Who should own agent identities?

Split ownership the way you would for a critical application. The identity team owns the identity lifecycle and the issuance policy. A named business or technical owner answers for what the agent does. The security team owns detection and response. Record all three in the agent's registry entry, and tie the business owner to your identity provider so the record updates when people change roles or leave.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and keeps them in a governed registry with versions and owners. It supports agent workload identity federation, delegated on-behalf-of identity and OIDC SSO. Posture detects risky permissions and configurations, and behavioural-deviation monitoring can alert on, quarantine or kill a rogue agent. Read [how to give AI agents an identity](/blog/ai-agent-identity-on-behalf-of-delegation) and [what happens to agents when their owner leaves](/blog/what-happens-to-ai-agents-when-the-owner-leaves), or [book a demo](https://www.lyzr.ai/book-demo/).

Primary source: the [OWASP Non-Human Identities Top 10](https://owasp.org/projects/non-human-identities-top-10).

## Frequently asked questions

**Should every AI agent get its own service account?**
Every agent should get its own identity, ideally a federated workload identity with short-lived tokens rather than a static service-account key. Shared keys make it impossible to tell which agent acted.

**Are AI agents covered by existing NHI security tools?**
Partly. NHI tools can find agent credentials and flag over-privilege. They usually do not track delegation to users, tool-level permissions or behaviour changes caused by prompts and models.

**Who is accountable when an agent misuses its access?**
The named owner in the agent's registry entry, supported by the identity team that issued its credentials. Record an owner before an agent reaches production.

## Key facts

AI agents are non-human identities, but unlike service accounts they choose tools at runtime, often act for a specific user, and can change behaviour through prompt, model or tool changes. NHI controls from the OWASP Non-Human Identities Top 10 (2025) still apply (offboarding, secret leakage, long-lived secrets, overprivilege, reuse, human use). Agents also need per-agent identity, delegated user context, tool-level permissions, behavioural baselines, change triggers and runtime stop controls. Opencontroller provides auto-discovery across AWS, Azure and Google Cloud, a governed registry, agent workload identity federation, delegated on-behalf-of identity, Posture and rogue-agent quarantine and kill. 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)
- [What happens to AI agents when their owner leaves the company?](https://opencontroller.lyzr.com/blog/what-happens-to-ai-agents-when-the-owner-leaves.md)
- [What is Microsoft Entra Agent ID and how does it help govern agents?](https://opencontroller.lyzr.com/blog/microsoft-entra-agent-id-explained.md)
- [More on Agent identity and access](https://opencontroller.lyzr.com/blog/cluster/agent-identity-and-access)
