On 20 August 2026 Microsoft published CVE-2026-69836, a deserialization flaw in Entra ID with a CVSS 3.1 base score of 10.0 and no patch for anyone to install. The FAQ on the record says Microsoft already mitigated the vulnerability and there is no action for users of the service to take 1. The next day Microsoft shipped revision 1.1, which flipped the exploited field from yes to no and described the edit as an informational change 1.

Both revisions sit in the table on that page, one line apart. I kept a copy of it, because the item that reaches my desk as “10.0, exploited, identity service” is a different item from the one I ended up filing.

The score measures a capability

The vector underneath the 10.0 reads AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H: reachable over the network, low complexity, no privileges required, no user interaction, scope change, high impact on confidentiality, integrity and availability 1. The string explains the number. The number rates a successful exploit.

Microsoft publishes a second score on the same page. The temporal score comes to 8.7, because exploit code maturity is Unproven and remediation level is Official Fix 1. Microsoft lists both, and the copies that reach me carry only the first. The 10.0 answers how much a working exploit would cost me. It says nothing about whether a working exploit exists.

NVD has not added its own assessment. As of its 18 October 2026 modification, the entry shows Microsoft’s 10.0 as the CNA score and NVD’s own base score as not yet provided 2. The number in circulation is the vendor’s, which is the right number here, since the vendor also fixed the code and holds the only view into the service.

No customer action is a category

Microsoft started issuing CVEs for critical cloud service flaws in 2024, including ones no customer had to patch, and tags that class exclusively-hosted-service. The Security Update Guide added a column and the API a note type so customers could filter these out 3. NVD displays the tag on this record 2.

The tag changes the shape of the work for me. There is no version to compare, nothing enters a patching cycle, and no artifact from this would show up in a scan of my estate. Microsoft fixed the flaw inside its own tenancy, before the CVE existed, and nothing in my environment changes as a result 1.

That is the trade the cloud CVE program makes: transparency instead of silence, in exchange for a slow drip of maximum-severity items nobody can act on. I fill a queue with noise if I treat each one as work to do, so I sort by whether there is anything in my estate to change before I let the severity number set the order, and most of these sort low once I read the tag.

The correction answered a narrower question

Help Net Security ran the story on 21 August and amended it on 24 August with Microsoft’s confirmation that the flaw was not exploited 4. SC Media quoted the spokesperson: the company corrected exploited to no, the vulnerability was not exploited in the wild, informational change only 5.

Roman Sannikov’s comment in that SC Media piece is the one I keep. Microsoft corrected the exploitation claim, he said, but never said out loud that no further action was required because there was no evidence anyone was compromised, and he called the distance between those two statements either an oversight or a hedge 5. John Strand’s point sits next to it: identity flaws arrive in classes, so the detective controls around the affected interface outlast any single CVE 5.

My reading of the correction is narrow. Revision 1.1 reads, “This vulnerability was not exploited in the wild,” and the FAQ on the same page says Microsoft had fully mitigated the flaw with no action for users to take 1. Neither sentence says that no one was compromised. The revision settled whether attacks happened; it left open whether any tenant was touched.

The checks I ran anyway

A withdrawn exploitation claim lowers urgency. It does not answer the question leadership asks in November about whether anything touched the tenant in August.

I started with the revisions table, not the summary. Revision 1.1 is dated 21 August and changes one field, with the score and the vector untouched. Microsoft’s technical assessment stayed put and only the evidence claim moved, so I can also tell which version of the record a given aggregator is repeating. A copy taken before 21 August still reads yes.

I pulled sign-in logs for 20 and 21 August and looked for successful authentications from countries we do not staff, token issuance for accounts that should have been idle in a summer week, and failures that turned into a success from a different network within minutes. A remote code execution inside an identity provider does not have to leave a sign-in behind, which is the limit of this check and the reason I do not treat a clean result as a conclusion.

I pulled the directory audit log for the same two days: new application registrations, credentials added to existing applications, consent grants, changes to privileged role assignments. I checked the emergency access accounts and whether either authenticated from somewhere unexpected. Conditional Access came next, looking for a policy that sat in report-only or carried an exclusion through that window, since a flaw in the identity layer matters more when a gap in policy sits behind it.

Last, I looked at the applications that federate to the tenant. Anything trusting this identity provider inherits its risk, and a vendor-side fix does not reissue the tokens already sitting in somebody’s browser.

None of that proves a negative. It is the evidence I can produce when that question comes up in November, and it took an afternoon.

Where the item sits now

Nothing from this is open in my tracker. There was never a patch for me to chase, and Microsoft withdrew the exploitation claim the day after it published the CVE.

Reading the two revisions gave me a habit for the next one of these. When I brief a maximum-severity item now, I quote the revision number and its date next to the score, because the same page carries a base score of 10.0 and a temporal score of 8.7, two numbers that measure different things, and Microsoft corrected one field inside 24 hours. Two months on, the only thing I can still point at in this item is the log review for those two days, and that is the part of a vendor-side fix that stays mine.


  1. Microsoft. (2026, August 20). CVE-2026-69836: Microsoft Entra ID remote code execution vulnerability. Microsoft Security Response Center. https://msrc.microsoft.com/update-guide/vulnerability/CVE-2026-69836 ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎

  2. National Vulnerability Database. (2026, October 18). CVE-2026-69836 detail. National Institute of Standards and Technology. https://nvd.nist.gov/vuln/detail/CVE-2026-69836 ↩︎ ↩︎

  3. Olson, L. (2024, June 27). Toward greater transparency: Unveiling Cloud Service CVEs. Microsoft Security Response Center. https://www.microsoft.com/en-us/msrc/blog/2024/06/toward-greater-transparency-unveiling-cloud-service-cves ↩︎

  4. Markovic, S. (2026, August 21). Microsoft patches critical Entra ID vulnerability (CVE-2026-69836). Help Net Security. https://www.helpnetsecurity.com/2026/08/21/microsoft-entra-id-vulnerability-cve-2026-69836/ ↩︎

  5. Zurier, S. (2026, August 24). Microsoft says Entra ID identity software not exploited, revises CVE. SC Media. https://www.scworld.com/news/microsoft-says-entra-id-identity-software-not-exploited-revises-cve ↩︎ ↩︎ ↩︎