
Most incident-response plans end the same way: reset the password, reimage the machine, revoke the session, rotate the secret. Microsoft's own documentation and three recent write-ups show that each of those steps cuts off less than the runbook implies, because the artifacts an attacker holds (an issued certificate, a Kerberos ticket, a refresh token, an application credential) live outside the thing being reset. Penetration tests and readiness exercises should therefore stop asking only "can the foothold be reproduced?" and also ask "after we respond, what still works?"
TL;DR
- Microsoft's analysis of two Active Directory Certificate Services (AD CS) intrusions states that a password reset or reimage does not revoke an issued certificate, and that revoking a certificate does not invalidate tickets already issued.
- Entra revocation stops new tokens, but access tokens already issued run until they expire; Microsoft documents lifetimes of about an hour by default and up to 28 hours in sessions that use continuous access evaluation (CAE).
- ERNW's 2026 testing found Token Protection binds refresh tokens only, and that its device-platform condition comes from a User-Agent string the client controls.
- A live campaign against managed service providers (MSPs) used valid application credentials, which sign in without a user and so bypass MFA and user-focused Conditional Access.
- Add a "does revocation work?" phase to engagements: list the artifacts, revoke, then measure what survives.
Four artifacts that outlive the response step
Issued certificates and Kerberos tickets
Microsoft's October 30 research post describes two intrusions that began with Teams help-desk impersonation. One abused a certificate template (the ESC1 pattern) to obtain certificates for privileged accounts; the other relayed a domain controller's NTLM authentication to AD CS web enrollment (ESC8) and used the issued certificate to obtain a Kerberos ticket-granting ticket (TGT). The post is clear about containment: resetting the user's password or reimaging the first workstation does not revoke an issued certificate or invalidate tickets already issued, and revoking a certificate does not retroactively invalidate TGTs or service tickets. It treats certificate revocation and ticket review as separate actions.
The same post records a second gap. Automated attack disruption restricted the compromised identities, but one privileged account was excluded from automated response. That account stayed available and was later used for logons to a file server and for data exfiltration. Microsoft's advice is to treat containment exclusions as incident decisions with a compensating manual action that someone verifies.
How long a stolen ticket lives is a configuration question. Microsoft's archived Kerberos policy documentation lists a default maximum user-ticket (TGT) lifetime of 10 hours and a service-ticket lifetime of 600 minutes in the Default Domain Policy. The documentation I read does not say whether a domain controller re-checks certificate revocation for a TGT that has already been issued, so a tester should measure that in their own lab rather than assume it. For related Kerberos background, see our earlier piece on RC4 and Kerberoasting after the default change.
Access tokens after "revoke sessions"
Microsoft's emergency-revocation guidance says that after the admin steps a user cannot get new tokens for applications tied to Entra, but for applications that use access tokens, the user loses access only when the access token expires. It says access tokens last one hour by default. A separate Microsoft page on token lifetimes says the default is variable, between 60 and 90 minutes, so the sources do not agree on one number.
CAE narrows the gap for supporting apps: it reacts to events such as account disablement, a password change or reset, and an administrator revoking all refresh tokens, with propagation latency of up to 15 minutes. The trade-off is that CAE-aware sessions can carry tokens valid for up to 28 hours, and Microsoft says the configurable token lifetime policy is not honoured for them. Initial CAE support is focused on Exchange, Teams and SharePoint Online. A responder who reads "tokens revoked" as "access ended" is wrong for the window between revocation and expiry.
Refresh tokens and what Token Protection binds
Token Protection is a Conditional Access session control that aims to make stolen tokens unusable from another device. ERNW's October 5 testing, run with Microsoft's recommended policy (Teams, Windows platform condition, mobile and desktop clients), reports two limits. First, it binds refresh tokens only; access tokens are never device-bound, so one already issued stays usable on any device for its lifetime. Second, the device-platform condition is derived from the request's User-Agent string, which the client controls. In their table, requests presenting a Windows User-Agent were blocked, while a Linux User-Agent obtained a usable, non-device-bound refresh token for Teams, and a Teams web-client request was not blocked. Their error timestamps are from May 27, 2026. Whether Microsoft has changed this behaviour since is not stated in the post.
Microsoft's own page frames Token Protection as one layer in defence in depth. It lists native-app support as generally available on Windows, iOS and macOS, with browser support limited to preview for selected web apps. ERNW's recommendation is to pair it with policies that block unknown platforms and require device compliance.
ERNW's November 4 follow-up covers a related gap. Under the Family of Client IDs (FOCI) mechanism, a refresh token issued to one client can be redeemed with another family client. ERNW reproduced this: a token from a SharePoint client could not read mail, but redeemed with the Outlook Mobile client ID it could. The post reports that Microsoft's position, relayed from Secureworks researchers who reported it in 2021, is that FOCI is deliberate because the token cannot exceed the user's overall access. The relevant question for a tester is whether containment logic and detection treat the client ID as a boundary, since it is not one here.
Application credentials
Field Effect's November 4 write-up of an active MSP campaign (first observed November 1) describes attackers using valid application credentials belonging to victims' own applications to authenticate to Entra ID and Azure DevOps. Because applications sign in without a user, the post says this sidesteps MFA and user-focused Conditional Access. It reports that Conditional Access blocked every replay of users' refresh tokens observed on November 3, and that disabled and rotated applications were retried in fixed 11-second loops. It also notes that already-issued tokens remain valid until they expire, and that how the credentials were obtained remains under investigation. Our earlier post on multi-tenant Entra app secrets covers inventory and reach; here the concern is what happens when you revoke.
A test plan for "does revocation work?"
The following is a scoping checklist, not a procedure. Run it only in a tenant and domain you are authorised to test, ideally in a lab first. None of it was executed for this post.
| Artifact | Response action to test | What to measure |
|---|---|---|
| Issued certificate | Revoke and publish revocation data | Whether the certificate still authenticates, and whether already-issued tickets still work until expiry |
| Kerberos TGT and service tickets | Disable account, reset password | Time until a held ticket stops working versus the configured lifetimes |
| Entra access token | Revoke sessions, disable user | Which resources still accept an existing token, and for how long, with and without CAE |
| Refresh token | Revoke sessions | Whether redemption under another FOCI client still succeeds before revocation, and fails after |
| Token Protection policy | Report-only first, then enforce | Whether platform enforcement depends on a client-supplied User-Agent in your configuration |
| Application credential | Remove secrets, disable the service principal | Whether sign-in attempts continue, and whether any granted access persists |
| Excluded accounts | Review automated-response exclusions | Which privileged accounts are exempt, and who acts manually |
Report the result as a time-to-effective-revocation per artifact. That is a more useful deliverable for an incident-response owner than a list of footholds.
The PentStark view
This section is opinion. The common thread is that identity controls are tested at issuance (was MFA enforced, was the device compliant) far more often than at revocation. A control can pass every issuance test and still leave a long tail of valid artifacts once something goes wrong. I would argue that measuring the tail should be a standard part of an identity-focused assessment, reported as time-to-effective-revocation, and that exclusions from automated response deserve the same scrutiny as exclusions from MFA.
I would hedge in two places. The Token Protection findings come from one team's tests dated May 2026, and the platform may behave differently now. And the lab-measured lifetimes may differ from documentation, which is why the table asks for measurement instead of quoting numbers. The "revocation gap" framing across these sources is mine; none of the authors present it as a single finding.
What to do this week
Identify privileged accounts exempt from any automated containment and assign a named manual action. Find where CAE is active and where long-lived tokens apply. Put Token Protection in report-only mode alongside platform-blocking and device-compliance policies if you rely on it. Decide in advance who revokes certificates and who reviews Kerberos tickets, since the two are separate steps.
References
- Microsoft Security Blog: help-desk impersonation and AD CS abuse, ESC1 and ESC8 intrusion paths (October 30, 2026)
- ERNW Insinuator: Token theft in Microsoft Entra ID, part 3, Token Protection (October 5, 2026)
- ERNW Insinuator: Token theft in Microsoft Entra ID, part 4, OAuth 2.0 standard gaps (November 4, 2026)
- Field Effect: application identity abuse in MSP customer environments (November 4, 2026)
- Microsoft Learn: Token Protection in Conditional Access
- Microsoft Learn: Continuous access evaluation
- Microsoft Learn: Revoke user access in an emergency
- Microsoft Learn: Configurable token lifetimes
- Microsoft Learn: Maximum lifetime for user ticket (Kerberos policy)
Last verified: 2026-11-11
