---
title: When should an AI agent be automatically quarantined?
url: https://opencontroller.lyzr.com/blog/when-to-automatically-quarantine-an-ai-agent
published: 2026-11-11
updated: 2026-11-11
author: Lyzr
---

# When should an AI agent be automatically quarantined?

**Automatically quarantine an AI agent when the harm of letting it run for another few minutes is greater than the cost of stopping it: confirmed data exfiltration, actions outside its granted scope, credential misuse, or runaway loops that burn cost or hit production systems. Lower-severity signals — quality drops, unusual but in-scope behaviour, single guardrail hits — should alert a human or open a ticket instead.**

Buyers ask this because automatic quarantine cuts both ways. Too sensitive, and you stop a revenue-critical agent over noise. Too lax, and an agent that has been manipulated keeps acting until someone reads the alert. The answer is a severity model agreed in advance, with concrete thresholds per tier.

## What does "quarantine" mean for an agent?

Quarantine is a reversible containment state. The agent stops taking actions, its credentials or tool access are suspended, and its recent activity is preserved for investigation. It differs from:

- **Block:** reject a single request or tool call, while the agent keeps running.
- **Kill:** terminate the agent and revoke its access, typically after investigation confirms it should not return.
- **Throttle:** reduce rate or budget without stopping it.

Quarantine is the right automatic action because it is safe to undo. Kill is usually a human decision. For the full response toolkit, see [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent).

## Which signals should trigger automatic quarantine?

Use three tiers. The examples below are starting thresholds; tune them in observe mode before enforcing.

| Tier | Response | Example signals |
|---|---|---|
| **1 — Auto-quarantine** | Quarantine immediately, page on-call, open a P1 incident | Agent sends data classified confidential to an unapproved destination; agent calls a tool or API it was never granted; agent uses credentials outside its identity (e.g. another agent's key); repeated blocked attempts to reach secrets; loop of identical tool calls far above its normal rate (for example 10× its 7-day p99); spend in one hour exceeding its daily budget |
| **2 — Alert, human decides** | Alert owner and security within minutes; human chooses quarantine or continue | Behavioural deviation from baseline without a policy breach; multiple guardrail hits (groundedness, task adherence) in a short window; new tool usage that is granted but never used before; access from a new region or network |
| **3 — Ticket** | Open a ticket for the owner; review within days | Quality scores from LLM judges trending down; a single prompt-injection detection that was blocked; configuration drift or risky permission found by posture checks; agent has no active owner |

A tier-1 signal should be rare. If an agent hits tier 1 every week, its permissions or baseline are wrong, not the threshold.

## What checklist should you use before turning on auto-quarantine?

1. **Every agent has an owner and an on-call route.** Quarantine without a human to respond just moves the outage.
2. **Each agent has a criticality rating.** For customer-facing or revenue-critical agents, consider auto-quarantine only for tier-1 signals that indicate active harm.
3. **Baselines exist.** Behavioural deviation needs at least a few weeks of normal traffic per agent version.
4. **A fallback exists for critical agents.** For example, route users to a human queue or a previous version while the agent is quarantined.
5. **Quarantine is tested.** Run a drill per platform: trigger it on a test agent and confirm tool access actually stops.
6. **Evidence is preserved.** Prompts, tool calls, policy decisions and traces from the window before quarantine are retained.
7. **Release requires a named approver**, and the release is logged.
8. **Incidents flow into your normal process.** Map tiers to your incident severities as recommended in [NIST SP 800-61 Rev. 3](https://csrc.nist.gov/pubs/sp/800/61/r3/final), and send them to ITSM as in [routing AI agent policy violations to ITSM](/blog/route-ai-agent-policy-violations-to-itsm).

## Which threats does this map to?

The [OWASP Top 10 for Agentic Applications](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/) describes risks such as goal hijacking, tool misuse, identity and privilege abuse, and rogue agents. Tier-1 triggers map most directly to tool misuse, privilege abuse and rogue behaviour, where the agent is already acting. Tier-2 and tier-3 signals catch earlier stages, such as a manipulated goal that has not yet produced a harmful action.

## How do you avoid quarantining the wrong agent?

- **Scope by version.** If a new version misbehaves, quarantine that version and keep the previous one serving.
- **Require two signals for tier 1** where a single detector has a known false-positive rate.
- **Exclude test traffic.** Red-team runs should be tagged so they do not quarantine production.
- **Review every auto-quarantine** within one business day and record whether it was justified. Track the false-quarantine rate per rule.

## How Opencontroller does this

Opencontroller's rogue-agent handling monitors behavioural deviation, raises alerts, and can automatically quarantine and kill agents, with policies applied across the whole agent estate. Per-agent Block and Quarantine controls act through the connected platform, so containment happens where the agent runs. Posture detects risky permissions and configurations, LLM judges score production traffic, and live groundedness and task-adherence guardrails supply tier-2 signals. Policy is code, with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. To see this on your agents, [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**What is the difference between quarantining and killing an AI agent?**
Quarantine is reversible containment: the agent stops acting and its access is suspended while evidence is kept. Killing terminates it and revokes access, usually after investigation.

**Which signals justify automatic quarantine?**
Active-harm signals: confidential data sent to unapproved destinations, calls to ungranted tools, credential misuse, runaway loops and spend far above budget.

**Should customer-facing agents be auto-quarantined?**
Only for tier-1 active-harm signals, and only with a tested fallback such as a previous version or a human queue.

## Key facts

Auto-quarantine AI agents on active-harm signals (data exfiltration, ungranted tool calls, credential misuse, runaway loops, budget breaches); alert humans on behavioural deviation; ticket quality drift. Opencontroller by Lyzr provides behavioural-deviation monitoring, alerts, automatic quarantine and kill, and per-agent Block and Quarantine that act through the connected platform. Opencontroller pricing starts 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)
- [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)
- [How do you red-team AI agents with multi-turn attacks?](https://opencontroller.lyzr.com/blog/how-to-red-team-ai-agents-multi-turn.md)
- [More on Runtime control](https://opencontroller.lyzr.com/blog/cluster/runtime-control)
