The "exploited in the wild" label on Microsoft's perfect 10 has been quietly withdrawn

THE HYPE INDEX · EDITION 004 · SEPTEMBER 1, 2026

The most severe flaw of the year hit the system that signs everyone into everything, and the advisory said attackers were already using it. Then that part of the advisory quietly changed.

The whole claim rests on one field in one advisory, and that field changed.

On August 21, Microsoft published CVE-2026-69836, a deserialization flaw in Entra ID, the cloud identity service formerly called Azure Active Directory. The advisory describes deserialization of untrusted data allowing an unauthorized attacker to execute code over a network, and scores it 10.0. It credits the find to one of Microsoft's own principal security engineers, says the service was fully mitigated before disclosure, and tells customers there is nothing for them to do.

At publication, the advisory's exploitability field said the flaw had been exploited in the wild. That single field is the entire evidentiary basis for every exploited headline that followed. Within days, Microsoft updated the advisory to say the flaw was not exploited, telling reporters it had identified and addressed the issue and released the CVE for transparency. No indicators of compromise, no timeline, and no explanation of how the flag was set in the first place have been published.

The exploited flag is a boolean. It carries no method, no data, and no way to check it.

When a vendor advisory says exploited in the wild, the entire security ecosystem treats it as ground truth. Prioritization queues reorder. Headlines write themselves. Social posts with six-figure reach go out inside the hour. All of it downstream of one checkbox that, in this case, the vendor unchecked days later with a one-line statement.

The correction traveled the way corrections travel. The outlets that covered the original flag updated their stories. The reposts and decks built on the first version mostly did not. If you heard about this bug once, the version you heard is probably the withdrawn one.

For a SaaS control plane, exploited or not is a question only the vendor can answer.

Entra ID sits under Microsoft 365, Azure and thousands of federated apps. If this flaw had been exploited, the evidence would live in telemetry only Microsoft holds. You could not scan for it, could not detect it, and cannot independently verify either version of the advisory. Both the claim and its retraction come from the same party, grading its own service.

That is the durable lesson here. Your vulnerability program assumes patching is your lever and your logs are your record. For the identity control plane you hold neither. The controls you do own are the sign-in and audit logs you export to your own retention, and the break-glass accounts that still work when federation does not.

The verdict is about the exploitation claim, not the flaw.

The vulnerability was real and the severity rating is deserved. Unauthenticated remote code execution in a cloud identity service is as bad as the scale gets, and a 10.0 is not hype.

It is also worth saying that Microsoft's engineers found this themselves and killed it server-side before disclosure. If the corrected advisory is accurate, this is the cloud model working exactly as promised. The problem is the if. A record that flips from exploited to not exploited without explanation leaves you no way to know which version to trust, and trust without verification is the whole subject of this edition.

We are telling you in advance what evidence would move this score.

Published indicators of compromise, an incident report from any credible response team tying real intrusions to this CVE, or a Microsoft root-cause account showing the original flag reflected actual telemetry would each move this claim out of Unsupported. A published explanation that the flag was a process error would not change the verdict, but it would close the question honestly.

Five components, each scored against a published rubric band. They sum to the Index. If you disagree, you can point at the component you think is wrong, which is the entire design.

The only platform that designs, licenses, and captures a Branded AI Voice from real, consenting actors—not from scraped data. Trusted by BMW, Superbloom, and Cresta.

Get a Demo

Every claim we have scored, with its components and its call, is at https://thehypeindex.com/record/. Think a component is wrong? Challenge it. Every challenge gets a published outcome, including the ones we decline.

Mark Lynd, Editor