The finance manager thought she was on a call with her CFO. The name was right. The profile photo was right, scraped from the public LinkedIn, which took all of forty seconds. The only thing that was wrong was the tenant the account belonged to, and nobody in a meeting ever looks at that.
This is the attack that has been quietly working against Teams for a while now, and the one Microsoft is finally responding to with the impersonation-detection features it announced on 17 November. Before we get to what the new controls check, it’s worth sitting in the failure for a minute, because the features only make sense once you’ve felt where the hole was.
How the fake CFO got in
Federated chat is the default posture for most tenants. External access is on, open to all domains, because somebody turned it on in 2019 to talk to a partner and nobody has touched it since. That setting does something people forget: it lets a user in any other Teams tenant start a chat with your people, and, this is the part that bites, set their own display name.
Display name is not identity. It’s a label the sending tenant controls. An attacker spins up a trial Microsoft 365 tenant, creates a user called Jane Okonjo, CFO, drops in a matching photo, and messages your finance team from a domain like contoso-finance-grp.onmicrosoft.com. Teams renders the name and the picture at full size. The actual UPN and tenant sit behind a hover or a click that busy people never make.
From chat it escalates to a meeting. A one-to-one Teams call, or a meeting link forwarded into the thread. The impersonator joins, the real employee sees the familiar name in the roster, and the social-engineering script runs: urgent wire, confidential acquisition, don’t loop in anyone else yet. In the incident pattern that prompted all this, the spoofed account sat through more than one meeting before a human finally expanded the participant card and saw the external badge.
Nothing about that chain was a software vulnerability. Every component worked as designed. That’s what made it so good.
What the new detection actually watches
The November features add two things that didn’t exist inside the meeting experience before.
The first is suspicious-participant detection. Teams now evaluates participants joining a meeting and raises an alert when something doesn’t line up, an external or federated account whose display name closely matches an internal user, an unexpected participant in a meeting that was expected to be internal, that class of signal. The point is the mismatch between who the name says you are and what tenant you’re actually coming from. That’s precisely the gap the fake CFO walked through.
The second is the Facilitator agent scanning the meeting transcript. Facilitator is the Copilot-based meeting agent Microsoft already ships for notes and action items. It now also reads the running transcript for impersonation cues, someone claiming to be a person they’re not, pressure language consistent with a social-engineering attempt, and can surface a flag during the meeting rather than in a log you read next Tuesday.
Read that second one carefully, because the dependencies matter more than the capability. The Facilitator agent needs the meeting to be transcribed, which means transcription has to be enabled and running, and it needs the Copilot licensing that Facilitator lives behind. No transcript, no scan. A meeting where everyone keeps their mic muted and types in chat gives the agent nothing to read. And it operates on language, it reasons about what’s being said, not cryptographic proof of who’s speaking. It’s a smart analyst listening in, not an identity check.
What happens when it fires
When detection triggers, the model is alert and surface, not block and eject. Participants and organisers get a warning that a participant may not be who they appear to be; the event is logged so your SOC can see it after the fact. The human in the meeting still makes the call.
I want to be blunt about this because the marketing will blur it: this does not slam the door on the impersonator. It taps you on the shoulder and says are you sure? That’s genuinely useful, the finance manager who gets a banner reading “this external participant’s name matches an internal user” is far more likely to stop and verify than one staring at a clean roster. But an attacker who stays quiet, keeps their claims vague, and lets the victim do the talking gives the transcript scanner very little to catch. The strongest defence here is still a warned, suspicious human.
Which meetings are covered, and which aren’t
The detection is aimed at the join paths that carry impersonation risk: federated participants from other Teams tenants, guests, and anonymous join via a link. Those are the three ways an outsider ends up in your meeting, and all three can present a chosen display name.
The edges are where it gets thin. Anonymous and dial-in participants give the transcript scanner audio at best and a typed label at worst, there’s no tenant to compare against, so the suspicious-participant signal has less to work with. And the whole system is a meeting control. The original compromise in the pattern above happened in chat, before any meeting existed. If your finance team gets talked into a wire transfer over federated chat and never starts a call, none of this ever runs. That’s not a flaw Microsoft is hiding so much as a boundary worth saying out loud.
Where you configure it
This lives in the Teams admin center, under the meeting and external-access areas you already know, Users > External access for federation scope, and Meetings > Meeting policies for participant and anonymous-join behaviour. The impersonation controls are surfaced alongside those settings rather than in some new console. The exact toggle names are still settling as the feature rolls out, so verify in your own tenant rather than trusting a screenshot from a blog, including this one.
The honest first move isn’t configuring the new feature at all. It’s finding out how wide your federation door is currently propped open, because the detection matters most exactly where you’ve let the whole world in.
Audit your exposure before you trust the new net
This is read-only reconnaissance using the modern MicrosoftTeams module. It tells you whether open federation and anonymous join are on, the two settings that decide how much the new detection has to defend.
Now walk the meeting policies, anonymous join is the setting that lets a link-holder in with a self-chosen name:
If you decide to tighten federation afterwards, scoping to named domains instead of the whole internet, do it against a reviewed list, never straight off a live query, and dry-run first:
Drop the -WhatIf only after you’ve read the planned change and confirmed the domain list is the one you meant.
What the incident teaches, stated plainly
- Display name is not identity, and the new detection exists precisely because people treat it as such. The control helps because humans won’t stop doing this. Design around the human, not the ideal.
- Detection is alert-and-surface, not block-and-eject. It makes a warned user suspicious. It does not remove the attacker from the call. Train the finance team to verify out-of-band anyway.
- The Facilitator transcript scan only runs where there’s a transcript and a Copilot licence. Silent, muted, or chat-only impersonation gives it nothing. Don’t budget it as coverage you don’t have.
- The protection is a meeting control; the first compromise usually happens in chat. Lock down federation scope, that’s the setting that decides who can even start the conversation.
- Open federation plus anonymous join is the worst-case surface. Audit both today. The new net is thickest exactly where you’ve already narrowed the door.
The fake CFO worked because every layer did its job and no layer was watching the seam between them. Microsoft just put a camera on one seam. The other seams are still yours.
