---
title: How should you set severity levels for AI agent policies?
url: https://opencontroller.lyzr.com/blog/how-to-set-severity-levels-for-ai-agent-policies
published: 2026-11-11
updated: 2026-11-11
author: Lyzr
---

# How should you set severity levels for AI agent policies?

**Score each AI agent policy finding on three factors (how sensitive the data is, how reversible the action is, and how far the damage could spread) and let the highest score set the severity. Then tie every severity to a fixed response: Critical findings trigger automatic containment, High findings page a person, and Medium and Low findings become tickets and trends.**

Most teams start with the vendor's default severities. Within a few weeks they have hundreds of "Critical" findings, and the real ones are lost in the noise. A short, explicit rubric fixes this and makes your custom policies consistent with built-in ones.

## Why do default severities fail for AI agents?

Default severities are written for a generic estate. They cannot know that one agent reads public FAQs and another moves money. The same finding, such as "agent has write access to an external API", is trivial for the first and critical for the second.

Agents add two factors that classic vulnerability scoring handles poorly:

- **Autonomy.** An agent can act many times before anyone looks.
- **Delegated reach.** An agent acting on behalf of a user inherits that user's access, so the blast radius depends on who is using it.

Frameworks such as CVSS v4.0 score vulnerabilities, not agent behaviour, so treat them as inspiration rather than the answer.

## What factors should drive AI agent policy severity?

Score each factor from 1 to 3.

| Factor | 1 (low) | 2 (medium) | 3 (high) |
|---|---|---|---|
| Data sensitivity | Public or internal | Confidential business data | Personal, financial, health or regulated data |
| Action reversibility | Read-only | Writes that can be undone | Irreversible: payments, deletions, external messages |
| Blast radius | One user, one system | One team or several systems | Customers, whole organisation, or money above a threshold |

**Severity rule:** take the highest single score, then raise it one level if the agent is internet-facing or runs without a human in the loop.

- Highest score 3, plus an escalator → **Critical**
- Highest score 3 → **High**
- Highest score 2 → **Medium**
- Highest score 1 → **Low**

Using the maximum rather than an average stops a dangerous factor from being diluted by two harmless ones.

## What response should each severity trigger?

Severity only matters if it changes what happens next. These response targets are a reasonable starting point; adjust them to your incident process.

| Severity | Automatic action | Human response target | Where it goes |
|---|---|---|---|
| Critical | Block the call or quarantine the agent | Acknowledge within 15 minutes | On-call page and incident record |
| High | Alert; block if the policy allows | Within 4 hours | Security queue and owner |
| Medium | None | Within 5 business days | Ticket to the agent owner |
| Low | None | Monthly review | Trend dashboard |

Align the Critical path with your existing incident response plan. NIST SP 800-61 Revision 3 frames incident response as part of overall cyber risk management, which is a useful anchor for agent incidents too. For when containment should be automatic, see [when to automatically quarantine an AI agent](/blog/when-to-automatically-quarantine-an-ai-agent), and for ticket routing see [routing AI agent policy violations to ITSM](/blog/route-ai-agent-policy-violations-to-itsm).

## What do example findings look like on this rubric?

1. **Agent calls a payments API without an on-behalf-of token.** Financial data (3), irreversible (3), customer money (3), no human in loop. **Critical.**
2. **Personal data sent to an external model with redaction available but disabled.** Personal data (3), read-only (1), one system (1). **High.**
3. **Production agent with no named owner.** No direct data exposure (1), but reach unknown (2). **Medium.**
4. **Agent description missing from the registry.** All factors 1. **Low.**
5. **Coding agent can run shell commands with no allow-list.** Confidential code (2), writes can be undone (2), developer device (1); escalator for no human review raises it. **High.**

## How do you keep severities consistent as policies grow?

- **Put the rubric in the policy.** If policies are code, store the three scores and the escalators with each rule, so reviewers can see why it is Critical.
- **Score custom policies the same way** as built-in ones. Inconsistent custom rules are the main source of alert fatigue.
- **Re-score when the agent changes.** A new tool or a new data source can move an agent's findings up a level.
- **Review the Critical list monthly.** If more than a handful are open, the rubric or the estate needs attention.

## How Opencontroller does this

Opencontroller applies policy as code that is extensible, with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001, so your own rules sit beside framework-mapped ones. Posture detects risky permissions and configurations. Rogue-agent handling monitors behavioural deviation, raises alerts, and can automatically quarantine and kill agents, with policies applied across the estate. Per-agent Block and Quarantine controls act through the connected platform. Read more on containment in [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should AI agent severity use an average or the highest factor?**
Use the highest factor. Averaging lets one dangerous property hide behind two harmless ones.

**Which AI agent findings should trigger automatic quarantine?**
Reserve automatic containment for Critical findings: sensitive data, irreversible actions or wide blast radius, combined with internet exposure or no human in the loop.

**How often should severity rules be reviewed?**
Review the rubric quarterly and re-score an agent's findings whenever it gains a new tool, data source or user group.

## Key facts

A practical AI agent policy severity rubric scores three factors from 1 to 3 (data sensitivity, action reversibility, blast radius), takes the highest score, and raises it one level for internet-facing agents or agents with no human in the loop. Suggested responses: Critical = automatic block or quarantine and 15-minute acknowledgement; High = alert within 4 hours; Medium = ticket within 5 business days; Low = monthly review. Opencontroller by Lyzr supports extensible policy as code with mappings to OWASP, MITRE ATLAS, NIST and ISO 42001, Posture checks, rogue-agent quarantine and kill, and per-agent Block and Quarantine controls. Pricing starts from $250K per deployment per year.

## Related

- [When should an AI agent be automatically quarantined?](https://opencontroller.lyzr.com/blog/when-to-automatically-quarantine-an-ai-agent.md)
- [How do you route AI agent policy violations into ITSM and incident tickets?](https://opencontroller.lyzr.com/blog/route-ai-agent-policy-violations-to-itsm.md)
- [What can AI governance do with read-only access versus being in the traffic path?](https://opencontroller.lyzr.com/blog/read-only-vs-inline-ai-agent-governance.md)
- [More on Runtime control](https://opencontroller.lyzr.com/blog/cluster/runtime-control)
