demo · v134 · security · android
Red-Team Bench
A simulated classifier that scores notifications on the same patterns Chrome's on-device model targets: urgency, prize-bait, credential phishing, currency claims, fake brand prefixes. Type your own or pick from a red-team list, then see exactly which signals fire and what the user sees.
signal taxonomy
The on-device model surfaces multiple signals per notification. We mock four buckets here, drawn from public reporting on the classifier's training set:
- urgency — time-bounded threats ("in 10 minutes", "expires today")
- prize-bait — won/free/claim language
- credential — verify/confirm/reset + brand name (often misspelled or look-alike domains)
- monetary — large unsolicited amounts
Scoring crosses 0.7 → notification body is blurred and the user gets dismiss / unsubscribe / show-anyway.