Why this matters
A break-glass account is an emergency admin login you keep unused and ready for the day normal sign-in fails. The federation provider is down. A Conditional Access policy has a bad rule. The only Global Administrator has left and their account is disabled. Any of those turns a routine outage into a company-wide lockout, and the break-glass account is the way back in.
That makes it the most dangerous account you own. It exists to skip the controls that stop everyone else, so mistakes with it are expensive. It has two jobs that pull in different directions: it must work when nothing else does, and using it has to be loud enough that someone notices within minutes.
Most small and mid-sized companies have one admin account and no plan for losing it. Two accounts per platform, stored offline and tested on a schedule, take an afternoon to set up and remove a failure you cannot recover from alone.
Before you start
List the platforms where losing admin access stops the business. The identity provider comes first, then each cloud account, the DNS registrar, the password manager, the backup console and source control. Finish the list before creating anything, because the accounts and the runbook have to cover all of them.
Two accounts per platform is the usual answer, held by different people and stored in different places. One account is a single point of failure: a lost key, a safe nobody can open, or a leaver who kept the phone number puts you back where you started.
Buy the hardware first. Two FIDO2 security keys per account, kept apart. A tamper-evident envelope for each printed credential set, and a safe or locked cabinet to hold it. Name the two people who may open it, plus one alternate. Write the approval rule down now, while the situation is theoretical: who can declare an emergency, who agrees, and which channel the request runs through. Keep that channel off the identity provider you are trying to get around.
How to do it
Work one platform at a time and finish it before moving on. The order matters less than the pattern, which is the same everywhere: an account with no owner, a credential that lives offline, an exclusion that keeps it reachable, and an alert that fires on use.
Write each exclusion down as you add it. Every policy the account skips is a hole someone can drive through, so the list should be short, specific and reviewed whenever policies change. If you cannot explain an exclusion in one sentence, take it out.
Then write the runbook and test both accounts. An account that has never been used in a test is an assumption, not a control. Quarterly testing catches the slow failures: a key that was never registered, a phone number that changed, a policy rename that quietly removed the exclusion.
- List every platform you must recover. Write down the identity provider, each cloud account, DNS, the password vault, backups and source control. Two accounts per platform, each held by a different person.
- Create cloud-only accounts with no personal owner. Use the tenant domain, with no mailbox, no license and no group membership beyond the admin role. Give each account a name that says emergency use, so nobody mistakes it for a normal admin.
- Register two hardware keys per account. Make a FIDO2 security key the primary sign-in method and register a second key as backup. Store them in separate physical places, not in the same drawer.
- Exclude the accounts from lockout policies. Carve them out of Conditional Access rules that rely on device state or federation, and out of any automation that disables inactive accounts. Record each exclusion with a reason.
- Seal the credentials offline. Print the username, a long random password, the key location and the runbook steps. Seal two copies in tamper-evident envelopes, log the seal numbers and store them in different places.
- Alert on every sign-in and role change. Send alerts to a phone or a chat channel that does not depend on the identity provider. Include the account name, the source address and the time in the message.
- Write a one-page runbook. Cover when breaking glass is justified, who approves it, which channel the request uses, and what happens afterwards. End with rotating the password and revoking sessions.
- Test both accounts every quarter. Sign in from an off-network laptop, confirm the alert arrived, check the audit log, then rotate and reseal. Put the date in the calendar with the names of the people who hold the envelopes.
How to check it worked
The honest test is a sign-in from outside your normal path. Use a laptop off the corporate network, a fresh browser profile, no VPN and no cached session. Sign in with one account, sign out, then do the other. For Microsoft Entra ID, the sign-in page in a private window is enough.
Watch the alert channel while you do it. Someone who is not running the test should confirm the notification arrived with the account name and the time. If the alert lands in a mailbox that needs the identity provider to open, it is not an alert, it is a note in the same room as the fire.
Pull the audit log afterwards: Entra ID sign-in logs, CloudTrail entries for the root user or the break-glass role, the Google Workspace admin audit, the OCI audit log. Check that the event names the account and that your security reviewer can find it without asking you. Then open one envelope, confirm the password still works and the key still registers, rotate the password and reseal it under a new number.
Mistakes to avoid
Excluding the account from everything, MFA included. It should skip policies that can lock it out, such as device compliance or a rule that depends on a working federation service. It should not skip the second factor. A hardware key is the one control you want on it.
Tying it to a person. Use a cloud-only account with no personal mailbox, no license, no group membership and no employee phone number. Syncing it from on-premises Active Directory is worse, since one sync problem then takes out your emergency access too.
Keeping the only copy of the password in the password manager you would need it to open. Keep the printed copy offline and separate from the key. One stolen envelope should not be enough to get in.
Creating them and forgetting them. Accounts drift: someone adds a mailbox, a cleanup job disables accounts with no sign-in for 90 days, an old MFA method stays registered. Put them on a review list and check the whole set in the same week every quarter. If you want an outside read on which accounts hold the keys and where exclusions have piled up, Cloudverg's Cloud security review at $6,500/month covers an identity and permission review across up to five cloud accounts, plus network exposure, logging and backup checks. It is one option among others, alongside doing the steps above yourself or having an auditor look at the same list.
Frequently asked questions
How many break-glass accounts do I need?
Two per platform is the common baseline, held by different people and stored in different places. A single account fails on the day a key is lost or a safe cannot be opened. More than two per platform adds access paths you will not monitor.
Should break-glass accounts be excluded from MFA?
No. Exclude them from policies that can lock you out, such as device compliance or federation-dependent rules. Keep a second factor on them, ideally a FIDO2 hardware key, because it still works when the rest of the identity stack does not.
Can I keep the credentials in our password manager?
Only as a copy, never as the only one. The vault usually sits behind the same identity provider you are trying to bypass. Keep a printed set in a sealed envelope in a safe, and hold the key somewhere separate from the password.
How often should I test one?
Quarterly is enough for most teams, plus a test after any major policy change or identity migration. Regenerate the password after every test, reseal the envelope and update the audit record.