demo · v134

Abusive Notification Filter

Chrome on Android v134 runs an on-device LLM over incoming notification text, flags abusive patterns (scams, social-engineering, fake urgency), and warns the user before they tap. Nothing leaves the device. The simulator below applies a heuristic that approximates the on-device classifier so you can see what kinds of text get flagged and why.

Heads up The real classifier ships in Chrome Android only and runs over notifications you receive in production. To exercise the actual signal in dev, opt-in on Android. The simulator below runs entirely client-side using a pattern-match heuristic.
probing UA / Notification permission...

Try it

no classification yet

The code

# Conceptual — actual classifier is in Chrome Android only:
const score = onDeviceLLM.classifyAbusive(notificationText);
if (score > THRESHOLD) {
  warnUser(notificationText);   // browser-managed warning
} else {
  showNotification(notificationText);
}
# No notification text leaves the device.

see also