WhatsApp Begins Limited Beta of On-Device Scam Alert

WhatsApp began a limited beta of Scam Alert on August 12, 2026, an optional feature that uses an on-device machine learning model to flag suspected scam messages from non-contacts without sending message content to Meta. Meta's engineering announcement details technical controls, including a third-party transparency ledger and Cloudflare-signed manifests, to make targeted model delivery detectable, and the company invited security researchers to test the implementation.

WhatsApp Begins Limited Beta of On-Device Scam Alert WhatsApp began a limited beta of Scam Alert on August 12, 2026. The optional feature uses an on-device machine learning model to flag suspected scam messages from non-contacts without sending message content away for classification or automatically reporting it to Meta. The company also published technical controls for model delivery and invited security researchers to test the implementation. WhatsApp began a limited beta of Scam Alert on August 12, 2026. The optional feature uses an on-device machine learning model to warn users about suspected scam messages from non-contacts while preserving WhatsApp's end-to-end encryption design. Meta's engineering announcement says message content stays on the device for classification and is not automatically reported to WhatsApp, Meta, or another party. PCMag independently reported the limited beta and the decision to publish a technical overview before wider availability. Local classification and user control When a user enables Scam Alert, WhatsApp downloads the model to the device. The model evaluates incoming messages from non-contacts for conversational structures and language patterns associated with scams. If the model flags a likely scam, the recipient sees an in-chat warning that is not visible to the sender. The user can block or report the contact, continue the conversation, or mark the chat as trusted so it is not flagged again. Meta says a user who marks a chat as trusted can separately choose to share the last five received messages to help improve accuracy; that sharing is optional and distinct from the local classification process. Verifiable model delivery Meta says every model version must be recorded in a third-party append-only transparency ledger before deployment. A release manifest contains SHA-256 hashes for the model weights, tokenizer, and related assets. Cloudflare, rather than Meta, signs the manifest digest with Ed25519 keys. Before loading a model, a device verifies the signature, checks the ledger entry and freshness, downloads the assets, and compares their hashes with the manifest. Meta says the client refuses to load the model if any verification step fails. These controls are intended to make selective delivery of a different model to one user publicly discoverable. Aggregate measurement without message content Meta's design still measures whether warnings are useful, but it describes the collected signals as counts of warnings and subsequent user actions rather than message content or conversation-level data. Those counts are processed through a confidential federated analytics pipeline, aggregated across users, and protected with differential privacy before approximate totals become available to WhatsApp. For ML and security practitioners, the release separates three problems that are often blended together: local inference, integrity of model updates, and privacy-preserving operational measurement. Running the classifier on-device addresses the default content-upload path; signed manifests and a transparency ledger address update provenance; and aggregate telemetry addresses evaluation without exposing raw conversations. The limited beta does not yet establish real-world detection quality. False positives, multilingual performance, adversarially written scams, device-resource costs, and the verifiability of the broader analytics pipeline remain important areas for testing. Key Points - 1Scam Alert performs optional scam classification on the user's device and does not automatically send message content or detections to Meta. - 2A third-party transparency ledger, Cloudflare-signed manifests, and client-side hash checks are designed to make targeted model delivery detectable. - 3The beta still needs evidence on false positives, multilingual coverage, adversarial robustness, device costs, and the privacy guarantees of aggregate measurement. Scoring Rationale The beta is a notable implementation of on-device ML for scam detection in a large encrypted messaging environment. Its technical relevance extends beyond WhatsApp because it combines local inference with signed model artifacts, transparency logging, and privacy-preserving aggregate telemetry. Sources Primary source and supporting public references used for this report. Practice interview problems based on real data 1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with. Try 250 free problems /problems