v134 · security

LLM-powered on-device detection of abusive notifications on Android

This launch aims to hide the contents of notifications that are suspected to be abusive. The user will then have the options to dismiss, show the notification, or unsubscribe from the origin. This detection is to be done by an on-device model.

concepts

  1. Abusive Notification Filter

    Chrome on Android runs an on-device LLM over notification text to flag abusive patterns. Privacy-preserving anti-scam.

  2. Scam classifier wave

    The attack pattern that drove this feature: prize/virus/credential-phishing notifications. Fire a wave and watch the classifier replace bodies with a reveal/dismiss/unsubscribe flow.

  3. Red-Team Bench

    Type your own notification, see exactly which signals fire (urgency, prize-bait, credential, malware-bait, suspicious-tld) and what the user sees.

  4. Notification Signal Map

    Pick a sample notification or type your own; a heatmap shows which abuse-signal categories fire and at what intensity. The combined score determines whether the body is redacted and the Show / Dismiss / Unsubscribe UI appears.

why it shipped

This launch aims to hide the contents of notifications that are suspected to be abusive. The user will then have the options to dismiss, show the notification, or unsubscribe from the origin. This detection is to be done by an on-device model.

references