---
title: What can you do to a misbehaving AI agent short of killing it?
url: https://opencontroller.lyzr.com/blog/graduated-responses-before-killing-an-ai-agent
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# What can you do to a misbehaving AI agent short of killing it?

**Use a response ladder: alert, throttle, restrict tools, force read-only, quarantine, and only then kill. Each step removes a specific capability while keeping the rest of the agent working, so you can contain the risk without breaking every workflow that depends on it. Write down in advance what triggers each step and who can authorise it.**

A kill switch is necessary, but it is blunt. Killing a customer-service agent because one tool returned odd results stops every conversation it handles. Most incidents call for something smaller and faster.

## What does a graduated response ladder look like?

| Step | What changes | Example trigger | Who can authorise |
|---|---|---|---|
| 1. Alert | Nothing; a person is notified and watches | Eval score drops below threshold on sampled traffic | Automatic |
| 2. Throttle | Lower request or action rate; tighter token budget | Spend per hour at 3x the 7-day baseline | Automatic, owner informed |
| 3. Restrict tools | Remove one tool or MCP server; keep the rest | A single tool shows unusual call volume or errors | Agent owner or on-call |
| 4. Force read-only | All write, send and delete actions blocked | Agent attempts an action outside its scope statement | On-call or security |
| 5. Quarantine | Agent stops serving users; state and logs preserved for review | Confirmed policy violation or suspected compromise | Automatic for defined triggers, otherwise security |
| 6. Kill | Agent disabled, credentials revoked | Data exfiltration, unrecoverable behaviour or legal instruction | Security lead or incident commander |

The numbers in the trigger column are examples. Set your own from each agent's baseline.

## How do you decide which step to start at?

Start at the lowest step that removes the harm. Three questions settle it:

1. **Is harm happening now or only possible?** Possible harm justifies alert or throttle. Active harm justifies read-only or higher.
2. **Can the harm be undone?** Reversible actions (a draft, a ticket comment) tolerate a lower step. Irreversible ones (payments, deletions, external email) justify jumping straight to read-only or quarantine.
3. **Is the cause local or the whole agent?** A bad tool response points to restricting that tool. Drift across every task points to quarantine.

Moving up the ladder can be automatic. Moving down should need a person, a cause and, for steps 4 to 6, a re-test.

## What should be decided before the incident?

- **Per-agent defaults.** Record in the registry which steps apply to each agent. A read-only research agent has no step 4.
- **Authority.** Name who can authorise each step out of hours. Automation can act at steps 1, 2 and 5 for well-defined triggers.
- **Mechanism.** Know where each step is enforced: a gateway in the traffic path, the agent platform's own controls, or credential revocation. A step with no enforcement point is a wish.
- **User communication.** Decide what users see when an agent is throttled, read-only or quarantined. See [notify users when an AI agent is shut down](/blog/notify-users-when-an-ai-agent-is-shut-down).
- **Evidence.** Log every step change with time, trigger, approver and reason. That record feeds the incident review.

## How does this map to incident response standards?

NIST SP 800-61 Revision 3 (April 2025) aligns incident response with the NIST Cybersecurity Framework 2.0 functions, including Respond and Recover. A graduated ladder is a containment strategy in those terms: it limits damage while preserving evidence and service. OWASP's work on agentic AI security also highlights excessive agency and tool misuse, which steps 3 and 4 target directly.

## How Opencontroller does this

Opencontroller monitors agents for behavioural deviation and raises alerts. It supports automatic quarantine and kill for rogue agents, with policies applied across the estate, and per-agent Block and Quarantine controls that act through the connected platform. Policy as code in the LLM Gateway and Agent Gateway sits in the path to allow, route, redact or block traffic, which is where tool-level restrictions are enforced when traffic flows through it. Enforcement depth depends on the connector and traffic path. See [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent), [when to automatically quarantine an AI agent](/blog/when-to-automatically-quarantine-an-ai-agent) and [circuit breakers for AI agents](/blog/circuit-breakers-for-ai-agents).

## Frequently asked questions

**Is quarantine the same as kill?**
No. Quarantine stops the agent serving users but keeps its configuration, state and logs for investigation. Kill disables it and revokes credentials.

**Should every agent have every step?**
No. Map the steps to what the agent can do. An agent with no write tools does not need a read-only mode.

**How fast should automatic steps act?**
Within the window that matters for the harm. For spend or looping, minutes. For data exfiltration, as close to real time as your enforcement point allows.

[Book a demo](https://www.lyzr.ai/book-demo/) to see graduated controls on your own agent estate.

## Key facts

Graduated response ladder for AI agents: 1 alert, 2 throttle, 3 restrict tools, 4 force read-only, 5 quarantine (state and logs preserved), 6 kill (credentials revoked). Start at the lowest step that removes harm; escalate automatically, de-escalate with a person. NIST SP 800-61r3 (April 2025) aligns incident response to CSF 2.0. Opencontroller (by Lyzr): behavioural-deviation alerts, automatic quarantine and kill, per-agent Block and Quarantine via connected platform, policy as code in LLM and Agent Gateway. From $250K per deployment per year.

## Related

- [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)
- [When should an AI agent be automatically quarantined?](https://opencontroller.lyzr.com/blog/when-to-automatically-quarantine-an-ai-agent.md)
- [What is a circuit breaker for AI agents?](https://opencontroller.lyzr.com/blog/circuit-breakers-for-ai-agents.md)
- [More on Runtime control](https://opencontroller.lyzr.com/blog/cluster/runtime-control)
