On 5 November Microsoft published its Q3 2026 email threat report. It counts 7.3 billion phishing detections between July and September, with nearly 3 billion in September alone, a 41% rise on August 1. The figure that stayed with me is smaller: about 129 million malicious messages that spoofed the sender domains of the business services Microsoft tracks 1. The reason many of them still left the filter is a rule someone wrote earlier.

The volume sat flat, which is its own finding

Microsoft puts the spoofing volume between roughly 41 and 46 million messages a month from May through September 1. DocuSign was the most spoofed brand in September at 29% of that month’s total, and Adobe, SurveyMonkey and Barracuda each picked up toward the end of the quarter 1.

I read a flat line on a familiar lure as evidence the lure keeps working. Microsoft describes the mechanism: the message claims the service’s actual domain as its sender identity rather than a lookalike 1. The request itself is ordinary, the kind I get from real vendors in a normal week: sign a document, approve an invoice.

Authentication failed and the message still arrived

The paragraph I read twice concerns delivery: many of those messages were delivered because organization- or user-configured exceptions overrode security protections, even after the messages failed authentication checks and were identified as malicious 12. In September, overrides applied to 63% of the messages with a recorded security outcome, and tenant-level overrides made up most of that share 1.

Those results form a sequence. SPF or DKIM or DMARC failed. The filtering stack scored the message malicious. Then an exception set at the organization or user level took precedence over both results 1.

Microsoft pairs the number with a caveat I want to keep attached to it: the figures do not establish that recipients opened the messages or that accounts were compromised 1. So 63% measures how often an earlier decision reached a message that authentication and scoring had already rejected. Delivery is where the report’s data stops, and a click, if one happened, would show up in the mail and endpoint telemetry I work with rather than in this report.

The exception started as a reasonable request

I have not met the administrator who set out to weaken mail filtering. The request that creates these entries is ordinary. A vendor’s notice lands in quarantine, the person waiting on it has a deadline, and the sender domain is the only stable thing in a message that changes every week. So the domain goes on a list, the notices arrive, and nobody revisits the entry.

The entry records a domain and little else. The context that made it safe lives with the person who wrote it: it was one team’s invoice notices, it exists because of a specific mail flow, and the flow will change when the vendor moves to a new sending domain. Nobody revisits those details, the entry stays in place, and a spoof of that domain inherits the exemption.

Microsoft states it in the report’s overview: “overly broad exceptions intended to preserve legitimate workflows can also expose them to impersonation” 1.

What I ask before an exception keeps its place

When I review the list, the questions are short, and each answer changes the entry.

The first is which message started it. If nobody can name the message, the entry has no case to answer and nothing to test it against.

The second is what the exception bypasses. An exception that skips quarantine carries less risk than one that suppresses the authentication failure or the malicious verdict. I read the second kind first, because it keeps working when the sender is impersonated. Scope comes with it: domain alone, or domain plus a recipient set. A sender-domain exception with no recipient condition covers every mailbox.

The third is the review date and the owner. An entry with a name and a date on it is a decision; an entry with neither is a default that nobody chose. I ask for both, and an entry that has a name but no date gets one before it keeps its place.

The last question is what would take it off. If nobody can name that condition, the entry has no end, and the only thing left to review is how far its scope reaches, which is the field I would narrow first.

The list is testable, the click is not

Microsoft’s 129 million describes its fleet, and the 63% is a share of Microsoft’s detections that carried a recorded outcome 1. That share says nothing about my tenant. The mechanism carries over though: a tenant-level override applies once for the whole tenant, so a single entry can cover every mailbox at once.

The audit itself is small. Export the exceptions that bypass the authentication verdict, sort by age, then by how broad the sender scope is, and read the oldest broad ones first. During a spoofing wave I check that export alongside the detections, looking for anything I wrote earlier that gave the message a path around the control that caught it.

The report cannot tell me whether anyone clicked. My own logs can tell me the message arrived, and delivery is where the decision stopped being mine.

An exception assumes a workflow will not change, and workflows change anyway: a new sending domain, a new vendor, a migration. The date on the entry tells me when the assumption gets tested instead of inherited, and the export I pull for the audit sorts by exactly that date.


  1. Microsoft Threat Intelligence. (2026, November 5). Email threat landscape: Q3 2026 trends and insights. Microsoft Security Blog. https://www.microsoft.com/en-us/security/blog/2026/11/05/email-threat-landscape-q3-2026-trends-and-insights/ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎

  2. Parker, S. (2026, November 5). Microsoft reports 7.3 billion phishing threats in Q3 2026. Redmond Channel Partner. https://redmondmag.com/articles/2026/11/05/microsoft-reports-7,-d-,3-billion-phishing-threats.aspx ↩︎