AntiMalware•August 12, 2026•🇷🇺Translated from Russian

WhatsApp Begins Limited Beta Testing of On-Device Scam Alert to Detect Fraud While Preserving End-to-End Encryption

WhatsApp, owned by Meta, has launched limited beta testing of its Scam Alert feature designed to warn users about potentially fraudulent messages from strangers.

The protection relies on a machine learning model that runs directly on the smartphone. Conversation content is never transmitted to WhatsApp or Meta, preserving end-to-end encryption. After activation, the device downloads the model and examines dialogue structure along with linguistic signs of fraud.

When a message appears suspicious, the recipient sees a warning while the sender remains unaware. Users can then block the contact, file a report, dismiss the alert, or designate the chat as trusted. In the trusted case, the last five messages can optionally be shared to help improve the model.

To prevent Meta from deploying a customized analyzer to specific users, every model release is recorded in an immutable log. A manifest containing SHA-256 hashes is signed with Ed25519 keys controlled by Cloudflare. The smartphone verifies the signature, log entry, and file integrity before executing the model.

WhatsApp collects only anonymized counters of detections and subsequent user actions. The app also includes a Scam Alert Activity log showing check results and model versions.

The current release remains an early technical preview. Meanwhile, Signal has introduced automatic key verification that uses a cryptographically verifiable log and allows users to confirm no unauthorized intermediary is present. Independent audits are performed by Cloudflare and Trail of Bits without access to user data.

Related articles

Habr•Fraud & Social Engineering

T-Bank's wev.js Anti-Fraud Collector: Reverse Engineering Browser Fingerprinting for Bot Detection

A detailed reverse engineering of T-Bank's wev.js anti-fraud SDK reveals how the bank builds a persistent device fingerprint during debit card applications. The collector draws an invisible 2000x200 canvas with non-existent fonts and colored arcs, then hashes the result with SHA256 to create a stable device identifier called nf. It also captures twelve data blocks including compressed mouse and keyboard event streams, full WebGL renderer details such as UNMASKED_RENDERER exposing the Apple M4 Max chip, and sensor data from gyroscopes and accelerometers. Eight dedicated bot-detection checks target automation frameworks by inspecting navigator.webdriver, native code descriptors, and Safari-specific markers. Experiments showed that spoofing the GPU renderer leaves the nf hash unchanged, while deterministic canvas modifications can create a stable fake device fingerprint. A second collector, fp-pixel.js, runs on SSO pages and aggregates even more signals including audio, fonts, and WebRTC before sending data to a separate endpoint.

AntiMalware•Fraud & Social Engineering

Scammers Impersonate Russian Traffic Police and Treasury Demanding Repeat Fine Payments

Russian scammers are sending SMS messages impersonating the State Traffic Inspectorate and the Federal Treasury, claiming that previous payments for traffic fines or state fees were not processed due to a system update. Recipients are instructed to pay the amount again or a small confirmation fee through a phishing link, with threats of accumulating debt if they refuse. The scheme was reported by the Moshilovka platform operated by the Narodny Front. Official agencies do not request repeated payments via third-party links or charge fees for transaction confirmations. Users are advised to verify payment status only through a verified Gosuslugi account and retain electronic receipts for any fees paid.

AntiMalware•Fraud & Social Engineering

Scammers Abuse Google Looker Studio to Send Fraudulent Compensation Emails

Kaspersky researchers have identified a new wave of Russian-language phishing emails distributed through the legitimate Google Looker Studio service, formerly known as Google Data Studio. The messages promise compensation payments or confirm supposedly approved money transfers from financial institutions, using the platform’s scheduled report delivery feature to reach victims. Emails arrive from a genuine Google address and carry the official Data Studio branding, making them harder for security tools to block. Recipients are directed to open documents in Google Presentations or forms in Google Forms that reference fake legal changes and display state symbols before leading to scam sites. The campaign is limited to messages sent at the start or middle of each hour due to scheduling constraints in the service. Experts note that the abuse of trusted platforms increases the difficulty of automated detection and filtering.

Habr•Fraud & Social Engineering

Human-like Cursor Movements: Evading Behavioral Anti-Fraud Detection in Browser Automation

Anti-detect browser provider Aurorium explains how modern anti-fraud systems analyze mouse telemetry including coordinates, timestamps, velocity profiles, acceleration, curvature, and USB polling intervals to distinguish humans from automated scripts. Standard straight-line movements at constant speed with perfect center clicks trigger immediate alerts because they lack natural human motor patterns described by Fitts's law. The article details four key techniques to simulate realistic behavior: cubic Bezier curves with random control points for organic arcs, phased acceleration and deceleration with occasional overshoot, low-amplitude Gaussian jitter that fades near targets, and offset clicks distributed around element centers. It also stresses that JavaScript click() calls set the isTrusted flag to false, making server-side mouse events via Playwright or Puppeteer mandatory. A complete Python implementation using Playwright's page.mouse API is provided, incorporating normal distribution sampling and smoothstep easing. The guide concludes that realistic mouse mechanics must be combined with clean residential proxies and consistent browser fingerprints to remain undetected.