Microsoft Entra Connect Health Vulnerability: Why It Went Undetected for 109 Days

The CyberSec Guru

Updated on:

Microsoft Entra Connect Health Vulnerability

If you like this post, then please share it:

Buy me A Coffee!

Support The CyberSec Guru’s Mission

🔐 Fuel the cybersecurity crusade by buying me a coffee! Why your support matters: Zero paywalls: Keep the main content 100% free for learners worldwide.

“Your coffee keeps the servers running and the knowledge flowing in our fight against cybercrime.”☕ Support My Work

Buy Me a Coffee Button

Last week Microsoft started notifying customers that a vulnerability in the Microsoft Entra Connect Health API had exposed tenant metadata and administrator contact details to unauthorized callers. The dates are the part that should hold your attention. The flaw was introduced on July 23, 2026. Microsoft detected it on November 9, 2026, which is 109 days later. A server-side fix shipped on November 19, ten days after detection and 119 days after the flaw went live. On November 20 the internal investigation closed with a confirmed finding: unauthorized parties had read tenant-level configuration data, though no configuration was changed.

Administrators first shared the notification on Reddit, and the security press picked it up over the weekend. It says: “On 9 November 2026, we detected a vulnerability in the Microsoft Entra Connect Health API that was introduced on 23 July 2026.” Microsoft says no passwords, password hashes, authentication tokens, valid cryptographic keys, on-premises topology or Azure subscription data were exposed. It also says only tenants that received a direct Azure Service Health advisory had confirmed unauthorized read activity. The total number of affected tenants has not been disclosed.

Some coverage shrugs this off as “just metadata,” and that reading misses two things. First, this is the second serious flaw in the same service this year. Between July and August 2026, Microsoft remediated a privilege escalation, disclosed by Cymulate Research Lab in September, that let a read-only Global Reader or Security Reader walk away with the live client secret of every Connect Health agent in a tenant. Second, the 109-day detection gap is itself the finding. A monitoring service whose whole job is telling customers when something is wrong ran a broken API for a third of a year without noticing, in a code path customers cannot observe. Microsoft fixed the code in ten days. It has not explained the 109 days, and a public root cause analysis is the deliverable that matters most.

What is Microsoft Entra Connect Health, and why does its API matter?

Microsoft Entra Connect Health is the cloud service that monitors hybrid identity infrastructure: the Entra Connect sync server, AD FS farms with their Web Application Proxy servers, and Active Directory domain controllers. It is the observability layer. It tells you whether replication is healthy, whether password hash sync is current, and whether your federation farm is throwing authentication errors. A lightweight agent on each monitored Windows server collects health and diagnostic data and uploads it over HTTPS to Microsoft, where it becomes the dashboards and alerts in the Entra admin center.

Per Microsoft’s installation documentation, the service requires an Entra ID P1 or P2 subscription, and the sync agent ships bundled with Entra Connect versions 2.5.79.0 and higher. Registering an agent takes a Global Administrator or Hybrid Identity Administrator account running Register-MicrosoftEntraConnectHealthAgent. The agent talks to endpoints including *.adhybridhealth.azure.com, *.servicebus.windows.net and *.blob.core.windows.net on port 443, and it updates itself.

Entra Health Connect
Entra Health Connect

How the Connect Health agent authenticates

The agent’s identity model is worth understanding precisely, because both of this year’s flaws live in its neighborhood. Cymulate’s research documents that each agent holds three registry values on its host: a tenant ID, a machine ID, and an AgentKey. The AgentKey is a 344-character client secret that Microsoft’s own software protects with DPAPI at rest. The tenant and machine GUIDs, joined by an underscore, form an OAuth client ID, and the AgentKey is the client secret. The agent authenticates with a client credentials grant to Connect Health’s own token service at s1.adhybridhealth.azure.com/oauth2/token. With the resulting bearer token it requests the two data-plane keys it needs: an Event Hub publisher SAS for streaming telemetry and a write-scoped blob SAS for uploading diagnostic data into Microsoft-hosted storage. In effect, the agent is a machine identity inside your identity infrastructure, condensed into a string the cloud honors.

The telemetry is more than operational decoration. For AD FS deployments, Connect Health processes audit events from every server in the farm to produce the Risky IP report and the failed sign-in reports, including extranet lockout breakdowns, which identity teams and SOCs use as a detection source. The service is also built to notice silence: if an agent cannot upload data for more than two hours, the portal raises a “Health Service data is not up to date” alert. Remember that alert. An attacker abusing the earlier flaw would trip it, and it is exactly the kind of alarm a mature operations culture learns to suppress as noise.

The GetTenant vulnerability: what data was exposed?

The November flaw was an information disclosure bug in GetTenant, the service call that returns tenant-level Connect Health configuration. According to the notification, the API returned too many fields to unauthorized callers:

  • the email addresses of the tenant’s technical contact and administrator accounts
  • the tenant ID, display name and initial domain
  • the country code and regional scope
  • the Microsoft Entra license tier
  • the directory synchronization status, including whether sync is enabled and the last sync timestamp
  • the tenant’s Connect Health onboarding state

The notification also says the flaw could potentially have allowed changes to Connect Health monitoring configuration, although the forensic review found only read activity. That potential write impact belongs in the record even though nobody exercised it. An attacker who can quietly edit monitoring configuration can blind the very service an incident responder would consult.

Sample Email Sent to Customers Informing about the Entra Health Connect Security Incident
Sample Email Sent to Customers Informing about the Entra Health Connect Security Incident

Is the leaked tenant metadata actually sensitive?

An honest assessment has to admit that some of these fields were never secret. While writing this, I fetched the unauthenticated OpenID Connect discovery document for a well-known tenant. It returned the tenant GUID in the issuer URL and a tenant_region_scope value, exactly as it does for any domain you query at login.microsoftonline.com/{domain}/v2.0/.well-known/openid-configuration. Tenant IDs, region scopes and initial domains are reconnaissance staples that attackers already harvest without touching a vulnerable API.

📬 Stay Ahead of Cyber Threats

Get the latest cybersecurity news, critical vulnerabilities, threat intelligence, tutorials, and exclusive giveaways delivered straight to your inbox. No spam. Unsubscribe anytime.

The genuinely non-public material is narrower and more interesting: the administrator and technical-contact email addresses, the license tier, the sync state and freshness, and the monitoring onboarding status. The field list reads like a projection of the tenant’s organization object, the same operational facts Graph exposes as dirSyncEnabled and lastDirSyncTime, joined with Connect Health’s own onboarding record.

How attackers can use the leaked data

Each non-public field lowers the cost of a follow-on attack in a specific way.

Administrator emails are raw material for spear phishing. Technical contacts are the people who receive Microsoft’s service notifications, so an attacker now knows exactly who would find a follow-up email about “the Connect Health incident” plausible. The breach notification itself is a ready-made lure template, and Microsoft’s own guidance to affected tenants warns about phishing that references Entra or Connect Health, which quietly admits the risk.

The license tier tells an attacker which defenses you can even deploy. Privileged Identity Management and Identity Protection require P2, so a tenant showing P1 has no risk-based Conditional Access and no just-in-time role activation to defeat.

The sync status and last sync timestamp reveal hybrid posture and freshness. Entra Connect’s default scheduler runs every 30 minutes, and a stale timestamp can point to an outage or a server in staging mode. In those windows an on-premises password rotation or account disable may not have reached the cloud yet.

None of this is a breach on its own. All of it is targeting data that makes the next phish, password spray or timing-sensitive intrusion cheaper and more convincing.

The expired pksCertificate field

The pksCertificate field is the most diagnostic artifact in the whole disclosure. The API returned an expired certificate that Connect Health had once used to register agents with an internal service decommissioned in late 2024. Microsoft’s position, quoted in the notification, is that the certificate “poses no risk” because it was no longer in use, but it “had not been removed from Connect Health storage and continued to be included in the GetTenant API response.”

Technically, Microsoft is right that an expired certificate is not a credential. X.509 certificates are public documents, and only the private key, which was not exposed, would matter cryptographically. The subject and issuer names may leak internal PKI naming and retired service hostnames, which has modest reconnaissance value.

What the field proves about the pipeline matters more. A production API was serializing storage artifacts from a service that had been dead for roughly eighteen months. The response was not built from an explicit, reviewed contract. It was built from whatever the storage layer happened to contain. That is a data lifecycle failure, and it rhymes with a much worse one. The Cyber Safety Review Board’s 2024 review of the Storm-0558 intrusion found that a consumer signing key created in 2016 reached a corporate debugging environment inside a 2021 crash dump and was never rotated. Combined with a token validation flaw, it let attackers forge tokens that Exchange Online accepted, compromising mailboxes at 22 organizations and more than 500 individuals. An expired public certificate and a live signing key are very different severities. The discipline they test is the same: knowing what cryptographic material you hold, where it lives, and when it should stop existing.

The earlier Connect Health flaw: when “read-only” handed out live secrets

To see why the security community’s patience is thin, read what Cymulate Research Lab published on September 3, 2026, in a post titled “Read-Only Is a Lie.” Researchers Ilan Kalendarov, Ben Zamir and Elad Beber showed that a principal holding only the Global Reader directory role, with zero Azure RBAC assignments and no on-premises access, could issue a single authenticated GET to management.azure.com/providers/Microsoft.ADHybridHealthService/services/{serviceName}/servicemembers/{serviceMemberId}/credentials?api-version=2014-01-01 and receive the agent’s live AgentKey in the response.

The authorization action behind that endpoint is named credentials/read, and Azure’s built-in Reader logic matches the wildcard */read, so a secret-issuing operation had been filed under a read verb. Microsoft’s own conventions say this should never happen. Storage, Cosmos DB and Event Grid keys come from POST listKeys actions deliberately excluded from read-tier roles. Key Vault Reader explicitly cannot read secret contents. The Automation documentation describes agent registration keys as a high-privilege operation precisely because a low-privileged actor could abuse them. ADHybridHealthService was the outlier.

The authorization bypass behind the Cymulate finding

The second defect was subtler and arguably more important. The provider is tenant-level, with no subscription segment in its resource paths, and it does not consult Azure RBAC at all. Cymulate’s differential testing showed an Azure RBAC Reader, whose role definition literally contains the matching wildcard, getting a 403, while a Global Reader with no RBAC whatsoever got a 200 and the secret.

Authorization runs off the wids claim in the caller’s ARM token, which lists Entra directory roles, checked against a hand-maintained allowlist. Global Reader, Security Reader, Security Administrator, Hybrid Identity Administrator and Global Administrator were allowed. Directory Readers and Reports Reader were denied. Security Reader was the floor.

This directory-role-to-ARM bridge is invisible to the tooling most organizations use to audit privilege. An Azure RBAC access review reports nothing, and BloodHound does not model the edge. Cymulate found the same pattern in a second provider, microsoft.aadiam, which backs Entra diagnostic settings, with a different allowlist. That suggests these lists are curated per provider and not derived from any consistent model.

From read-only access to impersonating an on-premises server

From there the attack chain is mechanical. The enumeration endpoints, also reachable by the read-only role and lacking any per-object authorization, return the whole fleet: every service member’s machine ID, plus forest and domain names, naming-context distinguished names, the sync service account name, and the SQL backend behind Entra Connect.

With a machine ID and a stolen AgentKey, the attacker performs the client credentials grant at s1.adhybridhealth.azure.com/oauth2/token and becomes, in Microsoft’s eyes, that specific on-premises server. The impersonated agent then requests its EventHubPublisherKey and BlobUploadKey from the monitoring policies endpoints, which gives it authenticated write access into the Event Hub pipeline and the blob container that feed hybrid identity monitoring. Cymulate verified the writes were real using conditional PUTs that returned 409 Conflict.

The researchers were careful about the limits. The SAS tokens are write-only and tamper-resistant. The agent token is an opaque Connect Health token that does not pivot to Graph or ARM. And the sign-in log spoofing that Dr. Nestori Syynimaa’s AADInternals research demonstrated from the same agent keys in 2021 no longer reproduced in 2026, apparently because Microsoft hardened ingestion.

The cruelest detail is that every read of the credentials endpoint minted a fresh key, so a GET rotated the production agent’s secret as a side effect. The legitimate agent, still holding the old key, is locked out, and recovery requires an administrator to re-register it. The attacker can simply read again. A “read-only” auditor thus produces credential theft and a monitoring outage with one request, which breaks the most basic contract in HTTP: RFC 9110 defines GET as a safe method, meant to be essentially read-only.

Cymulate scored the chain 8.3 on CVSS 3.1. MSRC rated it Important, confirmed the behavior on July 15 after a July 6 report, and declared it resolved, with rollout in progress, on August 10, 2026.

Did the earlier fix cause the GetTenant flaw?

Put those dates next to the November disclosure. The GetTenant flaw was introduced on July 23, 2026, eight days after Microsoft confirmed the credentials/read defect and eighteen days before the fix was declared resolved. The public record does not say whether the July 23 change was part of that remediation work or unrelated, and asserting causation would be irresponsible. It is exactly the question a root cause analysis exists to answer. “We broke a second authorization boundary in the same service while fixing the first one” is a hypothesis Microsoft should confirm or rule out in writing.

Why a 109-day detection gap is the real security finding

The vulnerability class has a name and a well-known detection problem. OWASP’s API3:2023, Broken Object Property Level Authorization, merged the old Excessive Data Exposure and Mass Assignment categories because the root cause is the same: authorization enforced at the object boundary and not at the field boundary. An API that returns too many fields still returns 200 OK, still validates against a permissive schema, and still meets every latency and error-rate objective. Nothing about it looks like an incident to the availability monitoring that dominates production engineering. That is how it can sit for 109 days inside a service whose brand is literally monitoring.

Static analysis does not catch it either, because there is no tainted data flow to trace. The code is type-correct, and the bug lives in the contract between the storage model and the wire format. A reviewer looking at a diff that adds a property to a shared DTO, or drops a projection filter, sees a one-line change with no injection sink and no obvious security meaning. Generic dynamic scanning flags extra fields only when a schema oracle tells it what the response should contain, and most crawling-based tools lack one.

How to prevent API data exposure bugs like this

The engineering controls that catch this class are known and unglamorous:

  • Build responses from explicit allowlisted DTOs and never serialize persistence entities directly, so a new storage column cannot reach the wire without a deliberate mapping change.
  • Snapshot the response shape in contract tests and fail CI on additive drift as well as breaking changes. In disclosure bugs, the addition is the regression.
  • Check data classification annotations on storage fields against external API contracts at build time, so a field tagged as contact PII or key material cannot appear in a public response without a reviewed exemption.
  • Scan sampled production responses for classified field names and PII patterns, which turns a 109-day blind spot into an hours-long one.
  • Review every remaining endpoint’s property-level behavior after fixing a credential-issuing authorization flaw. This service had just done exactly that in August, which made it an obvious candidate.

Cymulate recommended that organizations verify the fix independently instead of assuming it worked. The same advice applies to the vendor’s own adjacent code.

How confident is Microsoft’s “no changes were made” conclusion?

The notification leaves a forensic confidence question open. Microsoft says its investigators “reviewed available telemetry” and found no unauthorized configuration changes, and the word “available” is doing quiet work in that sentence. The exposure window ran 119 days, so the strength of a negative finding depends on whether telemetry retention covers it end to end. Customer-side Entra audit logs default to 30-day retention. Microsoft’s internal retention for this service is not public, and the notice does not state coverage. None of that means the conclusion is wrong. It means a root cause analysis should publish the retention basis so customers can weigh the conclusion themselves.

Why customers could not have detected these reads

Customers could not have detected these reads even in principle. The vulnerable calls ran inside Microsoft’s service, against Microsoft’s telemetry, which is how Microsoft identified the unauthorized callers. ARM read operations do not appear in the Azure Activity Log, which records control-plane writes, and calls to Connect Health’s tenant configuration API are not customer-visible directory events in Entra audit logs.

The shared responsibility model has an information asymmetry built in at exactly this layer. The provider holds the only logs that matter, so its detection engineering is the only control, and no longer one control among many.

Microsoft launched the Secure Future Initiative in November 2023 and framed it as a measurable, multi-year overhaul of engineering practice. Its November 2025 progress report tracked 28 objectives, five of them nearing completion. Three years in, a 109-day dwell time for a field-level authorization flaw in a hybrid identity service is a legitimate data point against which to measure those commitments. It belongs in the next progress report, and not only in a per-tenant advisory that most of the industry will never see.

What defenders should do this week after the Entra Connect Health breach notice

The patched API needs nothing from you. The residual risk is that your administrators’ email addresses and your tenant’s operational profile are now known to someone who should not have them.

Verify the notice and check your tenant data

Start by verifying the notice the way Microsoft recommends. Sign in to the Azure portal directly instead of following links in email, and check Azure Service Health for the advisory, because a fake version of this exact notification is a ready-made phish.

Microsoft published read-only reference scripts on GitHub. GetTenant.ps1 shows the current tenant-level fields the API returns, and GetTenantPksCertificateDetails.ps1 decodes the certificate metadata in memory so you can confirm its expired state. They run against ARM with GET requests only and need Az.Accounts 2.17.0 or later. Run them so your team knows what your tenant’s record actually contains. While you are in there, confirm the technical contact and notification recipients are current, since those addresses are both a phishing target list and your channel for the next incident.

Protect the exposed administrators from phishing

Assume targeted social engineering will reference this incident, your tenant or Connect Health specifics, and brief the SOC, IT operations, global administrators and the technical contact accordingly. Enforce phishing-resistant MFA (FIDO2 keys, Windows Hello for Business or certificate-based authentication) through Conditional Access authentication strengths on privileged roles, and use PIM for just-in-time activation where licensing allows. MFA does not make the emails secret again, but it removes the payoff from the credential phish those emails enable.

Hunt for downstream activity from July 23 to November 19

Review Entra sign-in and audit logs and the Azure Activity Log across the July 23 to November 19 window for downstream effects, with honest expectations. You are looking for suspicious sign-ins against the exposed accounts, unexpected role activations, new OAuth consents and configuration changes. You will not find the vulnerable reads themselves, because those were never yours to see.

Reduce Connect Health agent risk from the earlier flaw

Because the earlier flaw touched the same service, Cymulate’s advice still applies:

  • Inventory every holder of Global Reader and Security Reader, including the service principals belonging to compliance scanners and MSSPs.
  • Treat unexplained agent re-registration, or an agent going silent, as a potential indicator and not routine noise. A silent agent surfaces as the “Health Service data is not up to date” alert after two hours.
  • If you have any reason to suspect agent credentials were read before the August fix, re-register the agents to rotate them.

The root cause analysis Microsoft owes its customers

Everything Microsoft has published so far describes what happened and what was touched. Nothing explains how a change that widened an API response reached production on July 23 and stayed there until November 9. That gap is not academic. A useful root cause analysis would:

  • identify the introducing change and its review and test trail
  • state whether it was connected to the credentials/read remediation then in flight
  • explain what property-level authorization was supposed to enforce and why it did not
  • say what finally fired on November 9, whether a scanner, an anomaly detection or a human
  • disclose the caller’s authorization context and the telemetry retention behind the “no changes were made” conclusion
  • commit to specific preventive controls, with dates, of the kind described above

Microsoft has the machinery to publish this. MSRC handled Cymulate’s report quickly and substantively, and the Secure Future Initiative progress reports exist to make security engineering legible to outsiders. The ten-day fix is what competent incident response looks like. The 109 days before it are an assurance gap, and customers, who cannot audit Microsoft’s serializers or read Microsoft’s logs, have exactly one instrument for closing it: the explanation Microsoft chooses to publish. Microsoft has already shipped the scripts. The RCA is what it still owes.

Buy me A Coffee!

Support The CyberSec Guru’s Mission

🔐 Fuel the cybersecurity crusade by buying me a coffee! Your contribution powers free tutorials, hands-on labs, and security resources.

Why your support matters:
  • Writeup Access: Get complete writeup access within 12 hours
  • Zero paywalls: Keep the main content 100% free for learners worldwide

Perks for one-time supporters:
☕️ $5: Shoutout in Buy Me a Coffee
🛡️ $8: Fast-track Access to Live Webinars
💻 $10: Vote on future tutorial topics + exclusive AMA access

“Your coffee keeps the servers running and the knowledge flowing in our fight against cybercrime.”☕ Support My Work

Buy Me a Coffee Button

If you like this post, then please share it:

News

Discover more from The CyberSec Guru

Subscribe to get the latest posts sent to your email!

Leave a Comment

This site uses Akismet to reduce spam. Learn how your comment data is processed.