---
title: Which shadow AI agents should you deal with first?
url: https://opencontroller.lyzr.com/blog/prioritise-shadow-ai-findings
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# Which shadow AI agents should you deal with first?

**Deal first with agents that combine sensitive data, write actions and a long-lived credential. Score every finding on five factors — data class, actions, external reach, users affected and credential type — from 0 to 3 each. Act on anything scoring 11 or more within a day, and batch the rest by score.**

The first discovery run is usually a shock. Hundreds of agents, keys, bots and automations appear at once. Teams stall because everything looks urgent. A simple, consistent score fixes that.

This guide sits under [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents).

## Why not just triage by platform or team?

Because risk does not follow the org chart. A harmless summariser and an agent that can issue refunds can live on the same platform, owned by the same team.

What makes an agent dangerous is what it can touch and what it can do. The [OWASP Top 10 for Agentic Applications](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/) reflects this: tool misuse and identity and privilege abuse are listed risks because agents act, not just answer.

## What is the five-factor scoring rubric?

Score each factor 0–3. Add them up for a total out of 15.

| Factor | 0 | 1 | 2 | 3 |
|---|---|---|---|---|
| **Data class** | Public | Internal | Confidential | Regulated (personal, financial, health) |
| **Actions** | None (answers only) | Reads systems | Writes to internal systems | Moves money, sends external messages, changes access |
| **External reach** | None | Calls approved vendors | Calls unknown APIs or MCP servers | Internet-exposed or accepts outside input |
| **Users affected** | One person | One team | One department | Customers or the whole company |
| **Credential type** | Federated / short-lived | Scoped OAuth token | Personal access token | Long-lived admin or shared API key |

Two rules stop the score hiding real problems:

1. **Any single 3 on actions plus a 3 on data** goes to the top band, whatever the total.
2. **No owner found** adds 2 points. An ownerless agent cannot be fixed quickly.

## How do the scores map to response times?

| Band | Score | Response |
|---|---|---|
| Critical | 11–15 | Contain within 24 hours: rotate the credential, restrict or quarantine |
| High | 8–10 | Owner assigned and access reviewed within 7 days |
| Medium | 4–7 | Register, assign owner, review within 30 days |
| Low | 0–3 | Register and include in the next routine review |

These times are a starting point. Set your own to match your incident and change processes, and write them down so the score is not argued case by case.

## What does a worked example look like?

Three illustrative findings from a first scan:

| Finding | Data | Actions | Reach | Users | Credential | Owner? | Total | Band |
|---|---|---|---|---|---|---|---|---|
| Meeting-notes bot in chat | 2 | 1 | 1 | 2 | 1 | Yes | 7 | Medium |
| Refund agent in a SaaS platform | 3 | 3 | 1 | 3 | 2 | Yes | 12 | Critical |
| Script calling an LLM API with a shared key | 1 | 0 | 1 | 1 | 3 | No (+2) | 8 | High |

The refund agent goes first. It touches customer payment data, moves money and uses a personal token. The shared-key script comes next, mainly because nobody owns it and the key is long-lived. The notes bot waits for routine review.

## Which signals help you score quickly?

You rarely have perfect data on day one. Use what discovery already gives you:

- **Permissions and scopes** on the agent's identity tell you actions and data.
- **Connected tools and MCP servers** tell you reach.
- **Audit logs** show users and call volume.
- **Credential metadata** (age, type, last rotation) tells you credential risk.
- **Registry ownership fields** tell you whether an owner exists.

Where a factor is unknown, score it 2. Unknown is a risk in itself, and the default pushes teams to find out.

## What do you do after scoring?

1. Contain the critical band first: rotate credentials, remove write scopes, or quarantine.
2. Assign an owner to every high and medium finding.
3. Decide per agent: approve, remediate, replace or retire. See [what to do with a shadow AI agent once you find it](/blog/what-to-do-with-a-shadow-ai-agent-once-you-find-it).
4. Feed the score into your formal tiers. See [how to risk-tier AI agents](/blog/how-to-risk-tier-ai-agents).
5. Re-score when an agent gains a new tool, data source or credential.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, and connects to platforms such as Salesforce, SAP, Databricks and HubSpot plus Claude Enterprise and OpenAI Enterprise workspaces. Findings land in a governed registry with owners, versions, tools and MCP servers. Posture flags risky permissions and configuration, and per-agent Block and Quarantine controls act through the connected platform. See the [control plane overview](/blog/what-is-an-ai-agent-control-plane) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should every shadow AI agent be shut down?**
No. Most are legitimate work that skipped a process. Contain the critical ones, then register and review the rest.

**What if we cannot find who owns an agent?**
Add points for missing ownership and contain it if it scores high. An agent nobody owns should not keep live write access.

**How often should we re-score?**
Whenever an agent gains a tool, data source or credential, and at each scheduled access review.

## Key facts

A practical way to triage shadow AI findings is a five-factor score (0-3 each): data class, actions, external reach, users affected, credential type. Totals of 11-15 are critical (contain within 24 hours), 8-10 high (7 days), 4-7 medium (30 days), 0-3 low. Add 2 points when no owner is found; score unknown factors as 2. OWASP Top 10 for Agentic Applications lists tool misuse and identity and privilege abuse as agent risks. Opencontroller (by Lyzr) discovers agents, models and API keys across AWS, Azure and Google Cloud, flags risky permissions with Posture, and offers per-agent Block and Quarantine. Pricing starts from $250K per deployment per year.

## 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)
- [What should you do with a shadow AI agent once you find it?](https://opencontroller.lyzr.com/blog/what-to-do-with-a-shadow-ai-agent-once-you-find-it.md)
- [How do you risk-tier AI agents?](https://opencontroller.lyzr.com/blog/how-to-risk-tier-ai-agents.md)
- [More on Discovery and shadow agents](https://opencontroller.lyzr.com/blog/cluster/discovery-and-shadow-agents)
