Teams is adding AI phishing detection for risky chats — does that replace Safe Links and Defender for Office 365?

You’re right to ask whether “AI phishing detection” is a new control stack or another warning banner on top of what you already buy. Marketing loves the former; Steady Security prefers the latter until Message Centre and Learn say otherwise.

Here’s what’s going on. Roadmap 575288 (Phishing Detection in Teams Messaging) is In development, with general availability targeted for November 2026 (Worldwide; Android, Desktop, iOS, Mac, Web). The MRC blurb is carefully narrow: AI-powered detection that uses conversation context to identify phishing and social engineering in higher-risk messaging scenarios, primarily external first-contact chats. Microsoft states the capability is used only for security protection, and users may see warnings when suspicious activity is detected; additional protection may apply in certain regions or scenarios. Treat that as intent until Learn publishes admin toggles and false-positive guidance.

It sits beside, not instead of, today’s layers. Learn already documents Malicious URL Protection (link warnings in chats, channels and meetings), Safe Links for Teams when Defender for Office 365 policies include Teams, ZAP for Teams for malware and high-confidence phishing in internal chats/channels, user reporting, and the Teams admin centre Security detections report (impersonation, malicious URL, weaponisable files). A related roadmap item, 576198 (Near-real-time detection of malicious QR codes), extends URL-style protection to QR destinations — currently listed for GCC / GCC High / DoD with December 2026 GA — so Worldwide tenants should not assume QR coverage from 575288 alone.

This is Schofield’s Third Law for chat: the easier it is for an unknown external party to open a first-contact thread, the easier it is for social engineering to ride that trust. Steady before scale keeps process and existing Defender/Teams protections in place while you pilot the new warnings.

Options that actually work:

1. Keep the stack; add the warning as a layer. External access allow-lists (or a clear request path), Safe Links / malicious URL protection, ZAP where licensed, and user reporting stay on. Best Steady default when 575288 arrives.
2. Pilot with high-risk personas in Targeted Release / early rings. Finance, executive assistants and the service desk see external first-contact phishing most often — brief them on what the new warning means and how to report false positives before org-wide chatter.
3. Tighten first-contact behaviour now. Accept/block discipline for unknown external chats, no credentials or payment changes over Teams, callback on a known number — the same process rule you should already have for deepfake and callback fraud.
4. Turn off Safe Links “because AI will catch it.” Fast — and how a QR code, a weaponisable file, or a compromised internal account walks past a feature aimed at external first-contact context.

12 Panes action / comms: Tell Security: 575288 is conversation-context phishing warnings for higher-risk messaging, mainly external first contact — not a licence to remove URL protection or Defender Teams controls. Tell GRC: Add 575288 (and 576198 for sovereign clouds) to the AI/security register with “pilot → Message Centre verify → user brief” as the path. Tell users: A Teams warning is a stop sign; we still never share passwords or change bank details in chat. Leadership: Detection helps; process and existing Safe Links/ZAP remain the floor.

Benefits / risks: Context-aware warnings can catch social engineering that a raw URL scan misses; GA dates move; admin configuration and regional caveats are still thin on Learn while status is In development; QR malicious-code detection may lag or differ by cloud — don’t brief Worldwide staff as if 576198 were already theirs.

I’d suggest writing a half-page Steady note this month: keep Safe Links and malicious URL protection on, list the November pilot cohort, and reuse your “never authorise money or MFA resets in Teams chat” rule. When 575288 lights up in your tenant, the warning becomes reinforcement — not the only thing between a first-contact stranger and a very expensive afternoon.

DIGITAL SKILL WORKS LTD / 12 PANES · You ask, we answer