top of page

Who Can Read Your Conditional Access Policies, and Would Your SOC Ever Know?

Writer: Derek Morgan
Derek Morgan
8 minutes ago
8 min read

On Sat Nov 7, 2026, we ran the same Conditional Access read twice in our lab tenant, with the same tool and test account, 10 minutes apart. With no directory role, Microsoft Graph refused it with 403 AccessDenied. With Security Reader assigned and a fresh sign-in, the same call returned all 27 policies.


Then we looked for the read. The Microsoft Entra audit log, the Entra sign-in log, and the Microsoft Purview unified audit log held 0 records of it. What they did record was a password sign-in to Microsoft Graph from a command-line tool.


Our lab behaved as designed. Both audit logs record changes, and neither lists a read of these policies as an activity. That matters because the policies name every exclusion, trusted location, report-only policy, and uncovered app in the tenant.


Table of cloud service discovery calls: one 403 AccessDenied, one 200 success after Security Reader role was added.
Access attempt logs for Cloud Service Discovery on November 7, 2026, show an initial unauthorized access at 17:29 with a 403 error, followed by a successful access with Security Reader role at 17:39.

What an account holding a reader role can see

Listing policies through GET /identity/conditionalAccess/policies needs the Policy.Read.All permission and, for a signed-in user, a supported role. Microsoft documents 5 roles: Security Reader, Global Reader, Security Administrator, Conditional Access Administrator, and Global Secure Access Administrator (Microsoft Graph, List policies). 4 of them carry Microsoft's PRIVILEGED label. Global Secure Access Administrator doesn't, and the label is still in preview (Privileged roles and permissions).


Our Nov 7 test showed both sides of that gate. Limits: 1 tenant, 1 date, 1 role tested.


Security Reader and Global Reader also read per-sign-in Conditional Access results, Microsoft Entra ID Protection risky users and risk detections, Defender alerts, named locations, PIM data, and audit logs. Global Reader also reads the authentication methods policy (Microsoft Entra built-in roles).


Put together, 1 account can read the policies, see which ones applied to its own sign-ins, and check whether ID Protection or the SOC has flagged it. That chain is our assessment from documented permissions; Microsoft doesn't describe it as an attack.


Published intrusions come close. Microsoft's October 2023 Octo Tempest report lists configuring trusted locations in Conditional Access as defense evasion (Microsoft Security Blog; MITRE T1556.009). In August 2025, Microsoft reported that Storm-0501 surveyed existing protections, was blocked by Conditional Access, and then satisfied a hybrid-joined device requirement (Microsoft Security Blog). No report we found says an actor read the policies through Graph or used a reader role.


From no role to a reader role

Say an attacker controls an ordinary account with no role. The account can't assign itself Security Reader or Global Reader: assigning a Microsoft Entra role takes at least Privileged Role Administrator (Assign Microsoft Entra roles), and so does creating a role-assignable group (Role-assignable groups). Every path up starts from something an administrator configured earlier, or from a person who grants the role.


A way up the tenant already built

Eligible in PIM. An account with no active role can be eligible for a reader role and self-activate it within seconds. An earlier strong sign-in can satisfy MFA on activation, and the ticket number is never validated (Configure role settings in PIM). Learn publishes no default settings for the reader roles.


Owner of a role-assignable group. An owner of a group holding Security Reader can add itself, and can add other owners for persistence (Default user permissions). For PIM for Groups, Microsoft warns that "assignments that can be activated without approval can leave you vulnerable" (Role-assignable groups).


An approver. Where PIM requires approval, the approver needs no role (same PIM page), and whether anyone checks the request depends on your process (our assessment).


An app it owns. Member users can register applications by default and manage credentials on apps they own. A secret added to an owned app already granted Policy.Read.All reads the policies with no role and no MFA. APT29 added credentials to OAuth applications in the SolarWinds compromise (MITRE T1098.001). Posture, backup, and SIEM tools that read Conditional Access hold this permission (our assessment).


Or skip the climb

The attacker can also take over an account that already holds the role: token theft (adversary-in-the-middle or device code phishing), a help desk reset (CISA AA23-320A documents Scattered Spider doing this), or a partner technician whose GDAP relationship includes a reader role. Our lab's Security Reader assignment was standing, so a takeover would bring the role with no activation event.


Dead ends. Consent fails: Policy.Read.All needs admin consent, and a delegated grant still needs a reader role. On legacy Azure AD Graph, Truesec (2023) showed an end user exporting the policies with no role, and SpecterOps (Feb 2026) reported it still working. In our tenant, both apps we tried were refused (1 tenant, 1 date).

Path

Precondition

Log footprint

Fix

Activate an eligible reader role

PIM eligibility

Add member to role completed (PIM activation)

Approval and phishing-resistant MFA on activation

Add itself to a role-assignable group

Owns the group

Add member to group, which rules keyed on Add member to role miss

Audit group owners

Add a credential to an owned app (T1098.001)

Owns an app granted Policy.Read.All

Update application - Certificates and secrets management

Inventory those apps and owners

Take over a role holder

Token theft or help desk reset

Sign-in anomalies; Reset password

Phishing-resistant MFA; block device code flow

Partner (GDAP) access

Partner relationship with a reader role

Service provider sign-ins; no new role assignment

Remove unused GDAP roles


Infographic titled The climb and the read showing ways to reach Reader role, with audit and Graph log notes in blue, black and yellow boxes.
Visual representation of transitioning from having no role to obtaining a Reader role within Microsoft Entra, detailing steps like adding secrets, activating roles, and group assignments. The diagram highlights where each action appears in Microsoft Entra and Microsoft Purview logs, with emphasis on audit and sign-in logs.

What SecOps sees, and what it misses

By design, the read leaves no record in the Microsoft Entra or Microsoft Purview audit logs. The Microsoft Entra audit activity reference lists add, update, and delete activities for Conditional Access, and its one directory read is Read BitLocker key. The Purview audit log activities reference lists no Conditional Access activity and no read.


Our lab matched. From 16:30 to 18:30 UTC on Nov 7, the Entra audit log held 41 events and Purview 31 records, 0 of them Conditional Access. A SOC would see the sign-in before the read: Microsoft Graph Command Line Tools to Microsoft Graph, a password (ROPC) sign-in at 17:39:23, with a Purview UserLoggedIn record at the same second. Nothing says what was read.


Microsoft Graph activity logs are the one first-party record of the read, with URI, app, user, IP address, and the caller's roles. They need Microsoft Entra ID P1 or P2 and a diagnostic setting, and entries can take up to 2 hours to arrive (Graph activity logs). Our lab had 0 Entra diagnostic settings, so the read exists in no log there.


The climb is easier to see. Paths that change the directory leave named audit activities (the table above). Microsoft's privileged account guidance links Sentinel templates for roles assigned outside PIM and for PIM settings changes. Token theft and partner (GDAP) access change nothing in the directory, so they leave only sign-in records. The SOC's chance is the step before the read.


Attackers get cover from 3 places:

  • Microsoft's named alerts look for bulk enumeration tools, and AzureHound doesn't collect Conditional Access policies (Unit 42). We found no official Sentinel rule for these reads.

  • First-party app IDs resemble normal admin tooling, and token reuse lands in grouped non-interactive sign-in rows.

  • Up to 2 hours of latency lets a read-then-act attacker finish before the read is queryable (our assessment).


Our query, not Microsoft's:

MicrosoftGraphActivityLogs
| where RequestMethod == "GET"
| where RequestUri has "/identity/conditionalAccess/policies"
| summarize Calls = count()
    by UserId, ServicePrincipalId, AppId, IPAddress, UserAgent, bin(TimeGenerated, 1h)


Exclude baselined admins and tooling apps, flag service principals with an unexpected Policy.Read.All grant, and pair it with alerts on the climb.


What reading the policies gives an attacker

Conditional Access is evaluated after the first factor (Network signals). An attacker with a password and the policy list can pick a sign-in no enforced policy matches, without failed attempts.

What the policies reveal

Microsoft's fix

Excluded users and groups (emergency access, service accounts)

Passkey or certificate-based authentication and sign-in alerts for emergency access; workload identity Conditional Access for service principals

Trusted and named locations

Prefer device controls over IP trust

Report-only policies

Switch validated policies to On

Legacy authentication coverage

Block legacy authentication for all users and all resources

Device platform conditions (based on an unverified user agent, per Microsoft)

Pair with Intune compliance; block unsupported platforms

Resource exclusions

A baseline MFA policy on all resources with no exclusions

Admin policies without phishing-resistant strength

Phishing-resistant authentication strength for admin roles


Policy knowledge also makes by-design bypasses easier to pick. Fabian Bader's November 2025 write-up covers a device registration path that location controls can't block, which MSRC confirmed as expected behavior, and a Microsoft Authenticator bypass of the compliant-device requirement, fixed in April 2025 (Cloudbrothers).


The design rule: write Conditional Access assuming the attacker has read every policy.


The business case

Reading the policies replaces noisy trial and error with 1 quiet read: fewer failed sign-ins, fewer alerts. Quiet intrusions can run longer, and Mandiant's M-Trends 2026 puts global median dwell time at 14 days, up from 11. In our assessment, each extra day means more systems to clean up.


The controls have known costs. Graph activity logs need P1 or P2 plus about 15 GiB of Log Analytics ingestion a month per 1,000 users (Graph activity logs). PIM needs Microsoft Entra ID P2 or Microsoft Entra ID Governance (licensing). Authentication strength runs in Conditional Access, which needs P1 (Authentication strengths). The role review and app-owner inventory need staff time and no added license.


For executives: the climb to a reader role is logged and the read is not. Spend on the climb. Cut standing and eligible readers, require approval on activation, and alert when someone reads the policies.


Where Rejewski helps

Rejewski is our open-source, read-only toolkit (GitHub). It shows who can read your policies and whether your SOC would notice.


Today (v1.0.1):

  • tool_list_role_assignments and tool_list_pim_eligible_roles list standing and eligible holders of every directory role. In our lab, 2 principals were eligible for Global Administrator.

  • tool_list_ca_policies performs the read under a consent record and writes a timestamped replay-log entry. Run it, then ask the SOC to find the read by that time. Rejewski rates the finding high because policy knowledge helps an attacker evade defenses.


The Nov 7 check logged 35 entries, refusals included: a refusal is evidence that a control held.


Next (v1.1, in the backlog, not built yet): a no-role baseline run, findings that record the role each one needed, and read-only checks for the paths above. Today's tools already cover eligible reader roles.


Later (Stage 2, proposed in ADR 0009): emulating the app credential step itself, with verified rollback.


Checklist

  1. List every holder of a directory role, standing and eligible, starting with the 5 roles on Microsoft's List policies page. Include service accounts and GDAP partners.

  2. Put Security Reader and Global Reader under PIM, with approval and phishing-resistant MFA on activation. Check your own role settings.

  3. Audit owners of role-assignable groups, and alert on membership changes to those groups.

  4. Inventory apps granted Policy.Read.All and their owners. Set "Register applications" to No if users don't need it.

  5. Turn on Microsoft Graph activity logs, add the policy-read query, and alert on reader-role activations and app credential changes.

  6. Review your policies assuming they are public: no resource exclusions in the baseline, no reliance on location or platform alone, nothing important left in report-only, and exclusions backed by phishing-resistant credentials and sign-in alerts.


Conclusion

Every path to a reader role that changes the directory leaves an audit record; token theft and partner access leave sign-in records. The read itself, by design, leaves none in the Entra or Purview audit logs, and it reaches Graph activity logs only if you turned them on. Put your detection on the climb, and write your policies for an attacker who has already read them.

Tell us which attack stage Rejewski should emulate next in GitHub Discussion #148.

Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
bottom of page