- Microsoft says Teams will use AI to judge conversation context and flag phishing, including messages with no malicious link or file, starting in early November 2026.
- The feature focuses on external first-contact chats. Microsoft says admins need not enable it and that message content is not retained for model training or other secondary uses.
- Leaders should ask what is analysed, how flagged chats appear to users, and what is held for security purposes and for how long.
A message can be dangerous without containing anything dangerous
Most email and chat defences look for things. A bad link. A booby-trapped file. A known bad sender. Those checks catch much of what they are built for.
Some attacks give them nothing to scan. They open with a friendly message and build a little trust. Then they ask the victim to do the harmful thing themselves. The risk sits in the intent of the conversation, not in an attachment.
Microsoft is now building a defence aimed at that gap. The idea is simple. If the danger lives in the conversation, the defence has to read the conversation.
What Microsoft says it is adding to Teams
BleepingComputer reported on October 23 that Microsoft will start AI-powered phishing detection in Teams in early November 2026. The first phase is a targeted release expected to end by mid-November. BleepingComputer says worldwide availability is likely between early and mid-December.
Microsoft described the feature in a Microsoft 365 Message Center update. It said the tool detects "phishing, social engineering, impersonation attempts, and deceptive intent." It also said the tool looks for suspicious behavior that may lack "traditional indicators such as malicious links or files."
For most users, BleepingComputer says, the focus will be external first-contact chats. That means a person outside your organisation messages one of your staff for the first time. It is also the moment when a stranger can pose as a supplier, a colleague or IT support.
How it differs from the checks you already have
Older filters ask whether an item is known to be bad. This feature asks whether a conversation behaves like an attack. For example, a first message from a stranger might urge haste, claim authority or ask for a favour. A link scanner would see nothing wrong with it.
Those examples are ours, not Microsoft's detection criteria. The sources do not say how the model decides. They do not say which signals it weighs. They give no accuracy figures. Treat the approach as plausible and the performance as unproven.
The privacy promise, and what it rests on
Reading chat context raises an obvious worry. Microsoft addresses it directly. It says the detection is for security protection only. It lists productivity, summarization, advertising and user monitoring as uses it rules out.
Microsoft also says the detection system does not retain customer message content for model training or other secondary uses. That is a limit on purpose. It does not say that nothing is kept.
These are the vendor's commitments. The sources give no independent audit of them. In our view, commitments without independent verification will draw questions from staff, and leaders should expect to answer them.
Microsoft also says IT admins need not do anything to turn the protection on. That is convenient. It also means the feature may reach your users before your own review is done.
One more layer in a growing stack
BleepingComputer places this within a run of Teams security changes. Admins can lock external users through the Defender portal. A policy blocks identified external bots from meetings. Teams will blur QR codes from external senders. Users will be able to report suspicious guest invitations. Support for third-party deepfake detection in meetings was announced in early October.
Together, these show Teams being treated as a target in its own right. The lesson here is that chat has become a front door. A defence that only inspects objects leaves the conversation unwatched.
What to ask your team this week
Ask your security lead what exactly Teams will analyse. Ask whether external first-contact chats match where your staff are targeted. Ask how a flagged chat will look to an employee, and whether staff will know what to do with it.
Ask legal and privacy teams to check Microsoft's retention limits against your contracts. Ask what is held for the security purpose itself, and for how long. Ask them to check local rules on message analysis too. Ask for a plan to measure false alarms during the November rollout. Then you learn the real cost before December.
A tool that reads context is only as good as the person who reads its warning. The software can raise a flag. Someone still has to decide whether to stop.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Microsoft.
