Microsoft announced on 21 October that Teams will use AI-powered context-aware detection to identify phishing, social engineering, and impersonation attempts in external chats. The feature is entering targeted release in early November, with worldwide availability expected in early to mid-December 2026.
The move addresses a growing concern for organisations using Teams to communicate with external partners, suppliers, or students. Phishing attacks that exploit trust in familiar collaboration platforms have become more sophisticated, and education institutions—often handling sensitive student data and parental communications—are frequent targets.
What the new AI detection does
Microsoft's announcement describes a context-aware system that analyses message content in external chats to flag suspicious patterns indicative of phishing, impersonation, or social engineering. The company has emphasised that message content will not be retained for model training, addressing privacy concerns particularly relevant to regulated sectors like education.
External chats in Teams occur when users from different organisations communicate—for example, a school engaging with an external vendor or a tutoring centre collaborating with freelance tutors. These conversations sit outside an organisation's usual security perimeter, making them attractive entry points for attackers.
Why this matters for schools and tutoring centres
Education organisations often need to collaborate beyond their immediate team. Teachers may coordinate with supply agencies, tutoring centres may onboard contractors, and schools may engage external exam boards or support services. Each external conversation is a potential vector for credential theft, invoice fraud, or data breaches.
AI-driven phishing detection can help, but it isn't a silver bullet. Context-aware models still produce false positives and, more worryingly, false negatives. A phishing attempt that mimics legitimate communication styles or exploits time-sensitive scenarios—"urgent payment needed," "verify your login now"—can slip through.
For institutions already using Teams, the feature will roll out automatically. For those evaluating platforms, it's one more data point in a broader security conversation that should also include access controls, data residency, and how your platform handles external access in the first place.
The access model underneath
AI detection is a reactive layer. It analyses messages after they're sent. But phishing risk is shaped just as much by who can reach your organisation and how they get in.
Platforms built around shareable meeting links—whether in Teams, Zoom, or Google Meet—make external access frictionless. That's convenient, but it also means anyone with the link can attempt to join, whether they're phishing for credentials, recording the session without permission, or simply disrupting the class. Even with AI watching the chat, the damage may already be done.
Simpleclass takes a different approach: access is controlled by account and course enrolment, not a link. There is no meeting URL to forward or leak. Every participant is authenticated, and every session is tied to a specific course roster. External "guests" don't exist in the same sense—if someone isn't enrolled, they can't join, and phishing attempts via shared links become structurally impossible.
That design won't stop a determined attacker from targeting individual user credentials, but it removes an entire class of risk that AI detection is trying to solve downstream.
Privacy and data residency
Microsoft's assurance that message content will not be retained for model training is significant. Many education institutions—especially in the EU—are wary of how big-tech AI tools handle student and teacher data. Even anonymised or aggregated training data can raise GDPR compliance questions, and many schools have governance policies that simply forbid it.
Simpleclass hosts all video, recordings, files, and chat in the EU (Scaleway, Paris region) with a self-hosted media server. No data is sent to third-party AI providers, and there is no model training. For tutoring centres and language schools navigating data-protection obligations, that architectural choice often matters more than a feature announcement.
What to ask your platform
If you're using Teams and this feature is rolling out to your tenant, it's worth asking:
- How do we currently manage external chat access?
- Do we have governance policies that limit who can invite external users?
- Are our staff trained to recognise social engineering, even with AI assistance?
- Do we need external chat at all, or can we move sensitive collaboration into controlled, account-based environments?
If you're evaluating platforms—whether for live tutoring, exam prep, or language teaching—ask:
- Is access controlled by enrolment, or by shareable links?
- Where is session data (video, recordings, chat) hosted, and who can access it?
- Does the platform use third-party AI, and if so, what data is shared?
- What happens if a phishing attempt succeeds—can an attacker join live sessions, access recordings, or impersonate staff?
AI as one layer, not the whole story
Microsoft's phishing detection is a welcome addition to Teams' security toolkit, especially for organisations that rely on external collaboration. But security in online teaching is never just about one feature. It's about access control, data residency, transparency, and designing systems that make attacks structurally harder—not just algorithmically detectable.
If your institution is rethinking how it runs online classes, the question isn't just whether your platform can spot phishing. It's whether the platform's design reduces the attack surface in the first place.