SECURITY Signal 546
WhatsApp rolls out optional on-device scam alert that keeps messages encrypted and user-controlled
↗
Illustration only Photo by Alex Duffy on Unsplash
Engineers can deploy an on-device ML scam detection module that operates without breaking end-to-end encryption and gives users a toggleable warning.
Scam Alert provides a privacy-preserving way to surface potential scams while preserving the confidentiality guarantees of end-to-end encryption. By keeping all inference on the device and publishing model weights for independent verification, the feature lets security teams assess trustworthiness without exposing message content.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Scam Alert runs an on-device machine learning model that classifies incoming messages from non-contacts without sending any content off the phone.
Users must opt in to the feature, can turn it off at any time, and can mark chats as trusted to stop warnings for those conversations.
Telemetry is limited to differentially private aggregates processed in a confidential computing environment, and model versions are posted on a public transparency ledger before deployment.
THE READ
What the cluster adds up to.
WhatsApp introduced Scam Alert, an optional feature that runs a machine learning model entirely on the user's device to flag likely scam messages. The model processes incoming chats from people not in the user's contact list and produces a warning that only the recipient sees. All inference stays on the phone, preserving the end-to-end encryption of the conversation.
To use Scam Alert a user must first enable the setting, after which the app downloads the current model version to the device. The download happens once per model update and does not require ongoing server communication for classification. Because the model runs locally, the primary cost is the modest CPU and memory needed for inference, which WhatsApp states is compatible with typical mobile hardware. Users retain full control and can disable the feature or mark individual chats as trusted to suppress further alerts.
The system is limited to detecting patterns observed in previously reported scam conversations, so novel or highly tailored lures may evade detection. It only applies to messages from non-contacts, meaning scams that originate from a saved contact are not examined by the model. Improving accuracy relies on optional user actions, such as sharing the last five messages after marking a chat as trusted, which introduces a voluntary data-share step.
Transparency is built into the design: every model version is published on a public transparency ledger before it is deployed, preventing targeted delivery to specific users. Model weights are released so independent security researchers can verify that the model is purpose-built for scam detection only. Telemetry consists of differentially private aggregates processed within confidential virtual machines, allowing WhatsApp to measure feature usage without exposing individual data. These safeguards enable the broader security community to audit the implementation through the existing bug bounty program.
Written by elseif from the cluster below · checked for specifics the sources never containedTHE CLUSTER