Blog · Identity

The vendor's app secret is your standing access: testing multi-tenant Entra apps

PentStark TeamNovember 9, 20268 min read
Abstract circuit-trace cover in violet and cyan on a dark navy background

When a software vendor's multi-tenant Microsoft Entra app is compromised, the attacker does not need a single user account, password or MFA prompt in your tenant. They need the vendor's app credential, and whatever permissions your administrators once consented to. The October 2026 ConnectWise Cloud Backup incident is a live example. It also exposes a gap in how many tenant reviews and penetration tests are scoped: they check users, groups and admin roles, and stop before the third-party applications that hold standing, non-interactive access.

TL;DR

  • In a multi-tenant Entra app, the credential lives in the vendor's home tenant. Each customer holds only a service principal, and it carries the permissions that customer granted.
  • App-only sign-ins involve no user, device or MFA prompt, so the controls most teams lean on do not apply. Conditional Access for workload identities does not cover Microsoft or third-party multitenant apps.
  • Public reporting on the incident is consistent on mechanism and inconsistent on numbers, attribution and visibility. Read each figure for what it actually counts.
  • Tenant reviews and pentests should inventory consented third-party apps, the application permissions they hold, and who can add credentials, and should turn on Microsoft Graph activity logging before it is needed.

How the access works

Microsoft's documentation says an application has exactly one application object, which lives in the tenant where it was registered, the "home" tenant. A service principal is created in every tenant where the app is used, and for a multitenant app that happens when someone from that tenant consents. The service principal defines what the app can do in that specific tenant.

SpecterOps' explainer on app registrations and service principals walks through what follows from that split, using a hypothetical app and two lab tenants. The client secret exists only in the home tenant. To authenticate into a customer tenant, the caller presents the app ID and secret to the token endpoint with that customer's directory ID. Entra checks the secret against the home-tenant registration and issues a token carrying whatever the customer's service principal has been granted. The author's point, in the hypothetical, is that the customer cannot rotate a credential that does not exist in their directory. Their options are to disable the service principal or revoke its permissions.

Two consequences matter to testers. Consent does not expire, so a permission granted years ago is still live today. And an application permission such as Mail.Read is tenant-wide: it is a grant to read every mailbox, not one.

What the ConnectWise incident showed

SpecterOps' incident write-up reports that on 15 October 2026, between 06:13 and 08:34 UTC, the ConnectWise Cloud Backup application signed in to customer tenants from a single hosting-provider IP address. It quotes the key point plainly: MFA does not apply because this is an application sign-in, so no user, device or MFA prompt is involved. Huntress, which first reported the activity, says the authentication itself is certain, based on Microsoft's sign-in logs in each tenant.

The follow-on activity is the part worth studying. SpecterOps reports that in one tenant a second ConnectWise application granted itself Global Administrator and added new client secrets to two other trusted applications, which gives an attacker a way back in even after the original credential is reset. Huntress dates the self-assignments to 15 October and says two tenants showed post-compromise activity. Huntress also writes that it found no evidence of mailbox or file manipulation in the two tenants it reviewed, but cannot rule out reads.

How the vendor credential was obtained has not been disclosed, according to SpecterOps, and Huntress does not say either. We treat that as unknown.

Read the numbers for what they count

Three figures circulate and they measure different things.

  • Huntress' updates list 162 tenants with confirmed unauthorized sign-ins (as of 16 October), and later almost 1,000 tenants with sign-in activity from the named IP, of which over 300 succeeded. Huntress notes it can see sign-in logs for only some tenants.
  • Microsoft's threat intelligence post says the actor it tracks as Storm-0940, using a tool it calls ARTEX, attempted sign-ins to more than 2,250 tenants. That is attempts, not successes.
  • Microsoft also describes a separate password spray from 10 to 14 October, with more than 151,000 failed logons across 327 accounts in eight organisations, and states that no successful authentication, token issuance, Graph call or directory change was observed in it.

On attribution, Microsoft writes that it "assesses" Storm-0940 was responsible. In the text we read, the supporting link is infrastructure overlap: the spray source address matched one address in ConnectWise's advisory, and downstream sign-ins came from the same address range. No confidence level is stated. Huntress names no actor. The two accounts are not reconciled, and a reader should hold the attribution more loosely than the mechanism.

Why standard checks miss this

Three gaps recur across the sources.

Conditional Access does not reach vendor apps. Microsoft's documentation says policies for workload identities can be applied to single-tenant service principals registered in your tenant, and that Microsoft and third-party SaaS applications, including multitenant apps, are not covered. Creating or modifying such policies also requires Workload Identities Premium licences.

Visibility depends on what you enabled beforehand. Huntress says the Entra audit log does not record what an application read through Microsoft Graph, and that app-only actions in it carry no source IP. Microsoft's Graph activity logs are described as an audit trail of all HTTP requests the Graph service receives for a tenant, including the authentication method and service principal, but they must be streamed through a diagnostic setting and need Entra ID P1 or P2 plus a destination such as a Log Analytics workspace. The same page notes that it does not let you see a multitenant application's activity in another tenant. Service principal sign-in logs record non-user sign-ins using an app secret or certificate. Whether a given tenant had these enabled before 15 October is the question that determines how much it can reconstruct.

Graph tooling can hide the risky grants. SpecterOps' explainer reports that BloodHound draws no edges for read-only data permissions, so a service principal with no outbound attack path could still read every mailbox. Read that as a limit of graph-based path analysis, not a defect: it models control, not data access. (This and the next point come from a portion of the SpecterOps page that our fetch returned only as a summary, so verify them against the original.) The same piece reports that credentials added directly to a service principal do not appear in the admin centre, and advises checking the passwordCredentials and keyCredentials properties through Graph.

What to test

For an authorised assessment of your own tenant, or a client tenant in scope:

  1. Inventory service principals whose owning tenant is not yours, group them by vendor, and record the application permissions each holds. Prioritise tenant-wide read scopes and any scope that can grant roles, add credentials or modify applications.
  2. Compare each grant with what the product needs. Disable unused or unexplained apps rather than deleting them, so evidence is preserved, as both Huntress and Microsoft advise.
  3. Query credentials on every app and service principal you can see, including ones added after consent, and record who holds the rights to add credentials or assign roles.
  4. Confirm that service principal sign-in logs and Graph activity logs are on, where they go, and how long they are kept.
  5. Check whether detections exist for the patterns Huntress and Microsoft describe: credential additions to applications, an application assigning itself a privileged role, and a vendor app signing in from a new address.
  6. Record missing visibility as a finding or a scope limitation, never as a pass.

We have not run any tooling for this post, so the checks above are described, not demonstrated. Validate any query in your own environment first.

The PentStark view

This section is opinion. We think third-party application access should be scoped as an identity class in its own right, next to human accounts and admin roles, and that a pentest report which omits it is incomplete for any Microsoft 365 tenant. The incident does not show a new technique. Using a valid credential through its intended flow is the whole point, and that is why it slips past controls built for people. The public evidence for how the ConnectWise credential was taken, and for the strength of the attribution, is thin, and we would not build a programme on either. We would build it on the parts that are documented: the secret is not yours, the permissions are, and the logs only help if they were on. For how we approach scoping, see our methodology page. Related earlier posts: trusted publishing as an identity boundary covers the same theme of trust anchored to an identity you do not fully control, and RC4 and Kerberoasting after the default change covers another case where an identity-system default changed less than expected.

References

Last verified: 2026-11-09

Talk to an operator

Your next finding is one scoping call away.

Thirty minutes with a real operator tells us what you need and what we can deliver. No BDR handoff, no sales engineer theater. The person you talk to is the person who scopes the work.

● Responses in < 1 business day