Microsoft Teams adds AI phishing detection to external chats
Microsoft Teams will analyse conversation context to flag phishing, impersonation and deceptive intent in higher-risk external chats.
Microsoft is adding AI-powered phishing detection to Teams chats, with the first rollout aimed at spotting suspicious behaviour in conversations with external contacts. The feature is designed to recognise social engineering, impersonation and deceptive intent even when a message contains no malicious link or attachment.1
What Microsoft is changing
Microsoft says the protection analyses conversation context to identify potentially harmful behaviour. The initial focus is higher-risk situations such as a first conversation with an external sender, where an attacker may try to build trust or create urgency before asking for credentials or sensitive information.1
The feature is expected to begin a targeted rollout in early November 2026. Microsoft’s roadmap information says that phase should run through mid-November, with worldwide general availability expected between early and mid-December, although the company warns that roadmap dates can change.2
For most users, administrators will not need to turn on a separate setting for the protection. When Teams identifies a suspicious message, it can display a warning in the chat and give the recipient guidance on checking the sender’s identity and the request itself. Microsoft’s own example shows a message from a lookalike account named “Micros0ft Security” being flagged as a possible phishing attempt.1
That matters because conventional filters often look for obvious indicators such as a malicious URL or a file attachment. Microsoft says this detection is intended to find suspicious intent in the conversation itself, including messages that try to impersonate a trusted person or organisation without using those traditional signals.1
What happens to message data
Microsoft says the feature is being used for security protection only. It is not intended for conversation summaries, productivity insights, advertising or user monitoring, and Microsoft says customer message content is not retained for training its AI models or for other secondary uses.1
That promise is important because context-aware detection necessarily examines more than an isolated word or link. The value of the system depends on reading the relationship between messages, the apparent identity of the sender and the nature of the request. Microsoft’s published explanation presents that processing as a security measure rather than a general-purpose AI feature.1
The company is also adding other anti-abuse measures to Teams. Its roadmap lists protections for suspicious external chats and other changes intended to reduce fraud and phishing risks, while Microsoft has separately described controls for reported messages and malicious communications.2 The broader direction is clear: Teams is treating external conversations as a security boundary, not merely as another chat window.
Independent reporting from BleepingComputer says the feature will move through targeted release before a broader launch and quotes Microsoft describing the system as capable of identifying phishing, social engineering, impersonation attempts and deceptive intent.3 That report also places the change alongside a series of Teams safeguards aimed at external users and bots.
Our opinion
This is a sensible use of AI because it targets a problem that is difficult to reduce to a fixed blacklist. A scammer can avoid a suspicious domain, attach no file and still spend several messages pretending to be a colleague or supplier. Looking at the shape of that exchange gives Teams a chance to intervene before the victim reaches the dangerous part of the conversation.
The caveat is trust. Context-aware detection will need to avoid turning ordinary messages from new customers, contractors or community members into a wall of false alarms. Microsoft’s decision to frame the feature as security-only, and to say that message content is not retained for model training, is therefore as important as the detection itself. Users and administrators should still treat the warning as a prompt to verify a request, not as proof that every unflagged conversation is safe.
If the staged rollout performs well, this could make Teams less dependent on users recognising the visual tricks and urgency cues that modern phishing campaigns rely on. The real test will arrive in November, when the system meets messy, multilingual and genuinely legitimate external conversations at scale.