Microsoft Teams to flag suspicious chats using AI, not keywords

According to BleepingComputer, Microsoft Teams will start using AI-powered phishing detection in early November 2026. The feature looks at the context of a conversation, not just whether a message contains a bad link or a suspicious file.

That distinction matters. Most phishing detection today works by scanning for known indicators: a flagged domain, an attached file with a bad signature, a word pattern seen before. Microsoft says this new system instead analyzes how a conversation unfolds, which lets it catch impersonation and social engineering attempts that never contain a link or file at all.

Where it looks first

The rollout targets external first-contact chats, the moment someone outside an organization messages an employee for the first time. That is also where scammers posing as vendors, recruiters or IT support tend to start. A conversation that opens with urgency, a request for credentials, or pressure to act quickly can be flagged on tone and pattern alone, before anyone has clicked anything.

Microsoft says IT admins will not need to configure anything to turn the feature on. It also says message content will not be retained for AI training or any purpose beyond this specific security check. The targeted release phase is expected to run through mid-November, with general availability following worldwide between early and mid-December.

Part of a longer pattern

This sits alongside other Teams changes Microsoft has been rolling out. Admins have been able to lock out external users through the Defender portal since January, aimed at stopping ransomware groups from using Teams chats to approach employees directly. In August, a meeting protection policy began blocking external bots from joining calls automatically. More recently, Teams started blurring QR codes sent by external senders, a response to fraud attempts that hide a malicious link inside a scannable image. Users can now also report suspicious guest invitations directly inside Teams.

Each of these changes targets a different point where a stranger can first make contact: a chat message, a meeting invite, a QR code, a guest request. Taken together, they show Microsoft treating first contact itself as the moment worth watching, rather than waiting for a malicious link to appear later in the conversation.

This is the kind of shift Flggd tracks: detection moving earlier, from the artefact a scammer sends to the pattern of how they approach someone in the first place. For anyone using Teams at work, the practical change is quiet. No settings to change, no training to sit through. The protection is meant to work in the background, on the first message from someone a reader has never spoken to before.