WhatsApp Scam Alert Uses On-Device AI to Detect Fraud Without Breaking Encryption
WhatsApp has begun a limited beta rollout of Scam Alert, an optional feature that uses an on-device machine learning model to flag suspicious messages sent by unknown contacts. The classifier runs entirely on the recipient's device, meaning no message content leaves the user's phone and no automatic reports are sent to WhatsApp or its parent company Meta. When the model identifies conversational or linguistic patterns associated with known scams, the recipient alone sees an inline warning inside the chat; the sender is never notified. Users can then block the contact, report the message, ignore the alert, or mark the chat as trusted to suppress future warnings. Optionally, users who mark a chat as trusted may share the last five received messages to help refine the model's accuracy, a reminder that anyone exposed to credential-steering scams should periodically run an email breach checker to confirm their accounts have not already been compromised.
To prevent targeted model tampering, WhatsApp requires that every release be recorded on a third-party, append-only transparency ledger before distribution. Each release ships with a manifest of SHA-256 hashes covering the model weights and supporting files, and the manifest digest is signed using Ed25519 keys held by Cloudflare rather than Meta. Devices verify that signature, cross-check it against the ledger, and confirm the downloaded artifacts match the published hashes before the model is permitted to execute. Because classification stays local, WhatsApp built a confidential federated analytics pipeline that aggregates only two telemetry categories: the number of warnings triggered and the counts of subsequent user actions such as blocking or marking a chat as trusted. The company outlines a threat model covering external attackers, compromised infrastructure insiders, and supply-chain risk, arguing that singling out one user's data would require compromising the entire pipeline.
On the user-facing transparency side, recipients can review a Scam Alert Activity log under Account > Request Info, showing which messages were scanned, the outcome, and which model version produced the verdict. WhatsApp is also expanding its bug bounty program to cover Scam Alert, with researchers stress-testing the system ahead of the beta. Given the heavy reliance on device-resident inference, users concerned about their broader attack surface may benefit from a browser fingerprint test to understand what their device reveals online, and a full privacy checkup to audit exposed services. WhatsApp frames the release as an early technical preview rather than a finished product, with continued iteration planned throughout the beta based on researcher and user feedback before any general availability rollout.