---
title: Should AI agents ever get break-glass access?
url: https://opencontroller.lyzr.com/blog/break-glass-access-for-ai-agents
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# Should AI agents ever get break-glass access?

**No, not on their own. Break-glass access exists for humans in emergencies, and an agent that can elevate itself defeats the point. If an incident needs privileged action, a named person should hold the break-glass grant, approve any agent use of it, and the grant should expire in minutes or hours, followed by a review.**

This post covers why agents should not self-elevate, what a safe pattern looks like, and a review checklist. It is part of our [agent identity and access](/blog/ai-agent-identity-on-behalf-of-delegation) series.

## What is break-glass access?

Break-glass access is a pre-arranged way to get privileged access when normal paths fail: the identity provider is down, an admin is locked out, or an incident needs action faster than the approval queue allows.

Standards and vendor guidance treat it as a controlled exception:

- NIST SP 800-53 control AC-2(2) covers automated management of temporary and emergency accounts, including removing or disabling them after a set period.
- Microsoft's guidance for Entra ID recommends dedicated emergency access accounts, monitored sign-ins, and regular testing.

Both assume a human holds the credential and is accountable for its use.

## Why should agents not self-elevate?

Agents introduce risks that break-glass was never designed for.

- **Prompt injection can trigger it.** If an agent can request emergency rights, an attacker who controls its input can try to make it do so. OWASP lists excessive agency as a top risk for LLM applications for exactly this reason.
- **Agents misjudge emergencies.** An agent that sees errors may decide it is an incident and escalate, when the right move was to stop.
- **Accountability blurs.** "The agent used break-glass" is not an answer an auditor accepts. Someone has to own the decision.
- **Speed cuts both ways.** An agent with elevated rights can do a lot of damage in the minutes before anyone notices.

The safer default is the opposite: in an incident, agents lose access first. See [graduated responses before killing an agent](/blog/graduated-responses-before-killing-an-ai-agent).

## What is the safe pattern if an agent must help in an emergency?

Sometimes an agent is the fastest way to run a remediation script or summarise logs across systems. Use this pattern.

| Element | Safe design | Unsafe design |
|---|---|---|
| Who holds the grant | A named on-call human | The agent's own identity |
| How it is triggered | Human request with ticket reference | Agent self-request |
| Scope | Specific actions on specific resources | Admin role |
| Duration | Minutes to a few hours, auto-expiry | Until someone remembers |
| Agent use | Human approves each privileged action | Agent acts freely while elevated |
| Logging | Separate, high-severity audit stream | Mixed with normal logs |
| After | Mandatory review within days | None |

The pattern combines [just-in-time access](/blog/just-in-time-access-for-ai-agents) with human approval. The agent can propose a command; a person runs or approves it under their own break-glass identity.

## What should the post-incident review cover?

Use this checklist after any emergency elevation that involved an agent:

1. **Who triggered the grant**, when, and with which ticket or incident reference.
2. **Exact actions taken under elevation**, with the user and agent identity for each.
3. **Whether every privileged action was approved by a human** before it ran.
4. **Grant expiry.** Confirm it expired on time and was not extended without approval.
5. **Credential rotation.** Rotate any secret that was exposed during the incident.
6. **Agent state.** Confirm the agent's permissions returned to baseline, and record its version.
7. **Root cause.** Why normal access was not enough, and whether a standing runbook would avoid elevation next time.
8. **Policy updates.** Add any new block or alert rule the incident revealed.

Run a test of the break-glass path at least twice a year, just as you would [test the agent kill switch](/blog/test-your-ai-agent-kill-switch).

## What should be in policy as code?

Write the rules down where they are enforced, not only in a wiki:

- Deny any role-elevation request whose caller is an agent identity.
- Deny privileged actions by agents unless a human approval reference is present.
- Alert on any use of emergency accounts, with the calling identity.
- Auto-expire emergency grants and alert if expiry fails.

## How Opencontroller does this

Opencontroller supports policy as code across the agent estate, Posture checks that flag risky permissions and configurations, and per-agent Block and Quarantine controls that act through the connected platform. In an incident, that lets teams reduce an agent's reach quickly rather than expand it. Opencontroller does not provide human-approval hold workflows in its gateway; approvals for privileged actions should run through your existing identity and change tools. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Can an agent ever run a command under a break-glass account?**
Only if a named human approves that specific command and the grant is time-boxed. The human, not the agent, is accountable for the elevated action.

**Should agents be excluded from conditional access like emergency accounts are?**
No. Emergency human accounts are sometimes excluded so they work during outages. Agent identities should stay under every normal control.

**What is the first thing to do with agents during a security incident?**
Reduce their access: block or quarantine affected agents, revoke tokens, and preserve logs. Expanding agent access during an incident should be rare and human-approved.

## Key facts

Break-glass (emergency) access should be held by named humans, not AI agents. NIST SP 800-53 AC-2(2) covers automated management of temporary and emergency accounts. Safe agent patterns: no self-elevation, human approval per privileged action, scoped and auto-expiring grants, separate audit stream, and mandatory post-incident review. In incidents, agents should lose access first.

## Related

- [Can AI agents use just-in-time access?](https://opencontroller.lyzr.com/blog/just-in-time-access-for-ai-agents.md)
- [How do you revoke all of an AI agent's access at once?](https://opencontroller.lyzr.com/blog/revoke-all-access-for-an-ai-agent.md)
- [What does least privilege look like for an AI agent?](https://opencontroller.lyzr.com/blog/least-privilege-checklist-for-ai-agents.md)
- [More on Agent identity and access](https://opencontroller.lyzr.com/blog/cluster/agent-identity-and-access)
