Conditional access patterns that survive real users
Nathan Chadwick · · 9 min read · Updated
Listen
TLDRRead the short version
- Create two emergency access accounts, register a phishing-resistant method on each, and exclude those accounts from any policy that can block a sign-in.
- Roll policies in report-only first. Require MFA only after people can actually register, and block legacy authentication only after the sign-in logs show what still uses it.
- Named locations are for blocking countries you do not operate in. They are a poor reason to skip MFA from the office.
- Every exclusion needs an owner and a review date before you grant it. Review sign-in logs weekly for the first month after each change.
- Conditional Access is Entra ID P1. Risk-based policies need P2. Neither control covers a stolen session cookie.

The gate has to open for someone#
Conditional Access is a set of if-then rules on a sign-in. If this person, this app, this device, this location, then require MFA, require a compliant device, or block.
That description is tidy. The failure mode is not. The failure mode is a Global Administrator who cannot sign in to turn the policy off, because the policy applies to them, and the only other admin is on a flight. I have watched estates design the security and forget the person who has to undo it.
This post is the order I use so the rules get stricter without painting you into the portal. It is also the short list of things that usually break: a broad "require MFA" policy with no registration campaign, and a legacy-auth block switched on before anyone has read the logs.
What happens when more than one policy matches#
A sign-in is evaluated against every enabled policy whose conditions match. A few rules matter more than the rest of the UI:
- If any matching policy says block, the sign-in is blocked.
- If several policies grant access with requirements, the user has to satisfy all of them. MFA from one policy and a compliant device from another means both.
- Inside a single policy you can require all of the selected controls, or one of them. Across policies you do not get to pick. They stack.
- A policy in report-only does not enforce. It writes what it would have done into the sign-in log. Report-only policies do not need emergency-account exclusions, because they cannot lock you out.
"All cloud apps" includes the admin portals. A mistake there is a mistake you have to fix with an account the policy does not cover. Build that account before you enforce anything.
Microsoft's own common policies are the right starting catalogue: MFA for administrators, MFA for all users, block legacy authentication, and a tighter bar for privileged roles. I would rather run four policies I can name from memory than one policy that tries to be clever with fifteen conditions.
Emergency access, before the first enforce#
Microsoft's emergency access guidance is specific, and it is the first thing to do.
Create two or more cloud-only accounts. Not federated, not a daily admin, not a shared mailbox someone also reads. Register a phishing-resistant method on each: a passkey (FIDO2) is the recommendation, certificate-based authentication if you already run a PKI. Put the keys somewhere that is not the same safe as the password, and not in a ticket.
Then exclude those accounts from every Conditional Access policy that can block or restrict a sign-in. The passkey is what protects the account. The exclusion is what protects the tenant when a policy is wrong, or when the thing the policy depends on is down. An enforced policy on the emergency account can fail at the exact moment you created the account for.
A few habits keep the exclusion from becoming a back door:
- Exclude the accounts, not a group called
CA-Exclusionsthat grows every time something breaks. - Alert on any sign-in. A successful emergency-account sign-in is an incident until a person says otherwise.
- Sign in with them on a schedule and record that the sign-in worked. An untested break-glass account is a rumour.
- Keep Global Administrator on them. An emergency account that cannot reverse a Conditional Access change is not an emergency account.
Name the policy for the person who inherits it#
Six months later, nobody remembers why Policy-MFA-Final-v3 exists. A name that still makes sense has three parts: a number, who it hits, and what it requires.
CA010-AllUsers-AllApps-RequireMFA
CA020-AllUsers-AllApps-BlockLegacyAuth
CA030-AdminRoles-AllApps-PhishingResistantMFA
CA040-AdminRoles-AdminPortals-CompliantDevice
Put the owner, the exception, and the review date in the description. The description is the only place Entra will still be telling the truth when the ticket has been closed. Number them in the order you intend to enforce, so a new policy slotted in at CA015 is obviously a later idea.
The order that survives contact with users#
Plan it as a sequence. Each step has a log you read before the next one.
- Read a week of sign-in logs before you write a policy. You want to know which apps are real, which client apps are legacy, which countries you actually operate in, and which accounts are service accounts pretending to be users.
- Create the emergency accounts and confirm the alert fires when one of them signs in.
- Create named locations for a purpose. A country you do not operate in can be a block, or a reason to demand a stronger sign-in. The office range is not a reason to skip MFA. The old per-user MFA "trusted IPs" setting is a bypass, and it fights Conditional Access. Leave it off.
- Run a registration campaign before any require-MFA policy is on. Authentication methods policy, a deadline, and a person on the service desk who knows how to help someone with a new phone. If users cannot register, "require MFA" is "block the company".
- Put MFA for all users in report-only, scoped to a pilot that includes one awkward person and one admin path, then to everyone. Read the report-only results. "Not applied" and "user action required" are different outcomes. Fix the user-action pile before you enforce.
- Turn MFA on. Keep the emergency accounts excluded. Review the failures every day for the first week, then weekly.
- Block legacy authentication in report-only, then on, using the block legacy authentication policy. Filter the sign-in logs to client apps that are not modern auth. Scanners, postage machines, old line-of-business apps, and SMTP AUTH show up here. Give each one an owner. A dependency you have not named will become an outage on the morning you enforce.
- Move administrators to a phishing-resistant authentication strength. That built-in strength is Windows Hello for Business, a FIDO2 passkey, or multifactor certificate-based authentication. Authenticator push, SMS, and voice are not in it. Push is phishable. I covered why that matters in Identity risk is rising.
- Require a compliant or hybrid-joined device for admin portals, after Intune is actually marking those devices compliant. Do this after the Windows 11 rollout has a pilot that meets the compliance policy, not before.
- Add risk-based policies only if you have Entra ID P2. Identity Protection user-risk and sign-in-risk policies need P2. Business Premium and E3 give you Conditional Access. They do not give you risk-based Conditional Access. A policy you cannot license is not a control.
Session controls come last, and they should be boring. A shorter sign-in frequency on the admin portals is reasonable. A one-hour sign-in frequency on Outlook for everyone is how you train people to hate MFA and look for a workaround. I have written about that failure before: a control people cannot live with is a countdown to an exception, which is the point of the security quality-of-life post.
What usually breaks#
MFA with nowhere to register. The policy is on, the user has no method, the sign-in stops. The fix is the campaign in step 4, plus a time-boxed exclusion for the handful of people who are mid-replacement of a phone. The exclusion expires. It does not become a lifestyle.
Legacy authentication you forgot. Multifunction printers, backup software, PowerShell using an old module, a mailbox migration that finished in 2019 and left a connector behind. Block in report-only until the list is short enough to explain.
Guests. "All users" includes guests. A compliant-device requirement will block every partner on every sign-in. Exclude guests from the device policy and give them a narrower app policy, or accept that guests only use the apps you name. Do not discover this on the morning a supplier cannot open the project.
Workload identities. User policies do not apply to service principals. A script with a client secret is outside "MFA for all users". Workload identity Conditional Access is a separate kind of policy. If you do not have it, say so in the design, and keep the secrets in a vault with an owner.
The exclusion group. Someone adds themselves "until Friday" and Friday was eight months ago. I will not grant an exclusion that has no owner and no review date. Review the list monthly. If the reason is still true, write the new date. If it is not, remove it.
Device compliance in report-only on phones. Microsoft warns that a report-only compliant-device policy can still prompt people on macOS, iOS, and Android. Exclude those platforms from that report-only policy unless you mean to prompt them. A test that trains users to click through warnings is a bad test.
Report-only left on forever. Report-only is how you learn. It is not a control. Put a date in the description for when it either turns on or gets deleted.
What this will not catch#
Conditional Access decides whether a sign-in may start. It does not notice a session cookie that has already been stolen. Adversary-in-the-middle phishing completes MFA and walks off with the token. That is a detection problem, which is why ITDR sits next to this, not instead of it.
It also does not fix a shared local admin password, a helpdesk that resets MFA because someone sounded busy, or a privileged role that is permanently active. Those are the LAPS migration, the helpdesk process, and time-bound elevation. For a UK programme, the same MFA and admin-separation work shows up again in Cyber Essentials. Build it once.
A short checklist#
- Two emergency access accounts, each with a FIDO2 passkey, excluded from every blocking policy, alarmed, and tested.
- A week of sign-in logs read before the first enforce.
- Named locations used to block or step up, not to skip MFA.
- Registration campaign finished, then MFA in report-only, then MFA on.
- Legacy authentication blocked only after the remaining dependencies have owners.
- Phishing-resistant MFA for admin roles. Compliant device for admin portals once compliance is real.
- Risk-based policies only where Entra ID P2 is licensed.
- Every exclusion written down with an owner and a review date.
- Sign-in logs reviewed weekly for a month after each change.
Closing#
Conditional Access gets a reputation for exclusions because the first policy was too wide and the way back was a group. The way back should have been an account in a safe, a report-only week, and a log someone actually opened.
Start with the accounts that can unlock the gate. Then make the gate stricter in an order the service desk can explain. The policy that survives is the one a real user can pass on a Tuesday, and an attacker with a phished password cannot.
References#
Sources and further reading for the claims above.
- Microsoft Learn, Conditional Access overview
- Microsoft Learn, Plan a Conditional Access deployment
- Microsoft Learn, Common Conditional Access policies
- Microsoft Learn, Conditional Access report-only mode
- Microsoft Learn, Manage emergency access accounts
- Microsoft Learn, Authentication strengths
- Microsoft Learn, Block legacy authentication
- Microsoft Learn, Entra ID sign-in logs
- Microsoft Learn, Identity Protection risk policies (Entra ID P2)
- CISA, Implementing phishing-resistant MFA