---
title: What should you do with a shadow AI agent once you find it?
url: https://opencontroller.lyzr.com/blog/what-to-do-with-a-shadow-ai-agent-once-you-find-it
published: 2026-11-11
updated: 2026-11-11
author: Lyzr
---

# What should you do with a shadow AI agent once you find it?

**Don't delete a shadow AI agent on sight. Contain the risk first, find an owner, then decide within a fixed deadline whether to register it, block it or retire it.** Most shadow agents were built by people solving a real problem; a predictable triage process brings them under control without driving the next one further underground.

## What counts as a shadow AI agent?

A shadow agent is any agent, model deployment or model API key that is running but not in your approved registry. Common examples:

- an agent built in a SaaS platform's no-code builder by a business team,
- a cloud agent deployed from a developer's personal project or sandbox account,
- a model API key created outside procurement and embedded in a script,
- a coding agent or local MCP server on a developer laptop,
- a third-party tool with an embedded agent that was enabled by default.

Finding them is covered in [how to find every AI agent in your enterprise](/blog/how-to-find-shadow-ai-agents). This post is about the next step.

## What is the triage workflow?

Give every discovered item a status and move it through a small state machine. Each state has an owner and a deadline.

| State | Meaning | Exit criteria | Suggested deadline |
| --- | --- | --- | --- |
| Unregistered | Found by discovery, not in the registry | Initial risk check done | 1 business day |
| Contained | High-risk items restricted while reviewed | Owner identified | Immediate for critical findings |
| Under review | Owner found, information being collected | Decision made | 10 business days |
| Registered | Approved, with identity, owner and policies | Ongoing monitoring | — |
| Blocked | Not allowed to run; credentials revoked | Owner appeal or retirement | — |
| Retired | Removed, with evidence kept | — | — |

The deadlines are examples. What matters is that they exist and that an item cannot sit in "unregistered" forever.

## How do you decide what to contain immediately?

Run a quick risk check on every new item before anyone contacts the owner. Contain first if any of these is true:

1. **It has write access to a system of record**, such as finance, HR or customer databases.
2. **It handles regulated or personal data** without an approved route.
3. **It is exposed to the internet** or accepts input from outside the company.
4. **It uses a credential with broad scope**, such as an admin key or a long-lived personal token.
5. **It sends data to an unapproved model provider or region.**

Containment should be reversible: restrict permissions, rotate a key, or quarantine the agent so it cannot act while evidence is kept. Deleting it destroys the logs you need to understand what it did. See [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent) for the mechanics.

## How do you find the owner?

Start with metadata the platform already holds:

- the creator or last modifier of the agent or key,
- the cloud account, subscription or project and its owner tags,
- the identity that deployed it, matched to your identity provider,
- the cost centre billed for its usage.

If the creator has left the company, assign the agent to their manager or the owning team's lead. An agent without an accountable person should not stay in production.

## What does the owner need to provide?

Keep the review short so owners actually complete it:

- what the agent does and who uses it,
- which data and systems it touches,
- which models it calls,
- whether it acts on behalf of users,
- how it was tested,
- what breaks if it is switched off.

Much of this can be prefilled from discovery and logs. The owner confirms or corrects it.

## How do you decide: register, block or retire?

| Decision | When |
| --- | --- |
| Register | The use case is valid and the agent can meet policy, perhaps after changes such as a new identity, narrower tools or routing through the gateway |
| Block | The risk is unacceptable or the owner will not make required changes |
| Retire | Nobody uses it, a sanctioned agent does the same job, or the owner cannot be found |

Registering should give the agent its own workload identity rather than the creator's personal token. See [how to give AI agents an identity](/blog/ai-agent-identity-on-behalf-of-delegation).

## How do you stop shadow agents coming back?

Make the approved path faster than the shadow path: a self-service registration form, pre-approved model routes through a gateway, and standard identities for agents. Report the number of items found, contained and resolved each month, and the median time from discovery to decision.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and connected agent platforms, and Opencontroller Edge extends discovery to coding agents, skills and local MCP servers on developer devices. Discovered items can be brought into the governed registry with an owner and identity. Per-agent Block and Quarantine controls act through the connected platform, and Posture flags risky permissions and configurations. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should we notify the person who built a shadow agent?**
Yes, early and neutrally. Ask for help registering it rather than reporting a violation. People who feel punished stop telling you about new agents.

**Is a shadow API key as risky as a shadow agent?**
Often more so. A key can be used by many scripts and agents at once, and it usually has no owner recorded. Rotate unknown keys once owners are contacted.

**What if a shadow agent is business-critical?**
Contain the specific risk, such as narrowing permissions or routing it through the gateway, but keep it running while the owner registers it.

## Key facts

Shadow AI agent triage: unregistered → contained → under review → registered, blocked or retired. Contain first if the agent has write access to systems of record, handles regulated data, is internet-exposed, uses broad credentials, or sends data to unapproved providers. Opencontroller provides auto-discovery, a governed registry, and per-agent Block and Quarantine controls that act through the connected platform.

## Related

- [How to find every AI agent in your enterprise (including shadow agents)](https://opencontroller.lyzr.com/blog/how-to-find-shadow-ai-agents.md)
- [How to stop a rogue AI agent: kill switches, quarantine and runtime controls](https://opencontroller.lyzr.com/blog/how-to-stop-a-rogue-ai-agent.md)
- [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)
- [More on Discovery and shadow agents](https://opencontroller.lyzr.com/blog/cluster/discovery-and-shadow-agents)
