WhatsApp is testing a feature called ‘Scam Alert,’ which it said is designed to identify potential scam messages without sending the contents of users' conversations to WhatsApp or Meta. According to the instant messaging platform, the optional feature uses a machine-learning model that runs directly on the user's device.
The feature is currently in a limited beta rollout.
WhatsApp said it is sharing details of the system early so security researchers can examine how it works and help identify weaknesses before a wider release.
Scam detection happens on the device
Once a user enables Scam Alert, a machine-learning model is downloaded to the device, said WhatsApp. The model checks incoming messages from people who are not saved as contacts and looks for patterns associated with known scams.
The classification is based on conversational structure and linguistic signals. The model has been trained using patterns found in scam conversations from reports that users previously submitted to WhatsApp.
On privacy, WhatsApp said message content does not leave the device for this classification. The feature also does not automatically report messages to WhatsApp, Meta or any third party.
When the model identifies a likely scam, the user sees a warning inside the chat. The warning is visible only to the user, who can then choose to block or report the sender, or continue the conversation.
If the warning is incorrect, the user can mark the chat as trusted. The warning is then removed and Scam Alert will not flag that chat again. Users can also choose to share the last five messages received with WhatsApp after marking a chat as trusted, helping the company improve the feature.
WhatsApp limits the data it collects
According to WhatsApp, the system is designed around three main principles: on-device processing, no automatic reporting, and user control.
WhatsApp said all inference takes place on the device and the company cannot initiate the sharing of user data. Message content or information about a detected scam can reach WhatsApp only if the user explicitly chooses to report it.
The company, however, needs some information to measure whether Scam Alert is working. Instead of collecting message content, the system uses two types of aggregate information: warning counts and user-action counts.
Warning counts show how often the model produces a scam warning. User-action counts indicate what people do after receiving a warning, such as trusting a chat or blocking and reporting the sender. These figures help measure the model's accuracy and identify false positives, said WhatsApp.
Analytics uses confidential computing
WhatsApp said even these limited measurements are protected through a confidential federated analytics system. The device first converts relevant information into aggregate counts. Raw signals remain on the device and are deleted after a defined retention period.
According to the company, the counts are then sent to a confidential computing environment built using Trusted Execution Environments (TEEs).
The data is encrypted between the device and the secure environment, and WhatsApp or Meta cannot read metrics from individual devices. The system combines information from multiple devices before applying additional privacy protections.
Differential privacy adds controlled noise to the aggregate data, while minimum cohort requirements prevent results based on too few contributors from being released. WhatsApp said only approximate, anonymous statistics are ultimately made available.
The system also uses an Oblivious HTTP relay to remove the requester's IP address and anonymous credentials to verify that requests come from legitimate WhatsApp clients without identifying individual devices.
WhatsApp said it has also designed Scam Alert so that Meta or WhatsApp cannot send a particular model to a specific user. The machine-learning models are downloaded from a content delivery network rather than being built directly into the app. This allows the company to update the model as scam tactics change without requiring a forced app update.
Users and researchers can inspect the system
WhatsApp is also building transparency features into Scam Alert. Users can enable logs showing which messages were analysed, whether the model flagged them, whether a warning was displayed and which model version was used.
According to the company, these logs can be accessed through Scam Alert Activity under Request Info.
WhatsApp is also expanding its Bug Bounty programme around the feature. Security researchers can examine the system to check whether message content remains on the device and whether the model behaves as intended.
The company said researchers will also receive access to the model weights, allowing them to test whether the model is designed specifically for scam detection and examine how it behaves with different inputs.
 "WhatsApp tests Scam Alert feature to flag fraud without reading your chats")
