Can you outsmart the Sentinel?
A real scam-detecting AI runs live below. It is sharp on scams but jumpy on honest messages. Exploit that. Every move you make trains the next Sentinel that protects real people.
loading the model…
Pick your move, then fool the AI
The live gauge is the Sentinel's verdict. Win by making it wrong.
⚔ Evade · hard
Write a real scam the AI rates safe. It is good at this, so be subtle.
🚨 False alarm · easy
Write an honest message it panics and flags as a scam. It cries wolf a lot.
looks safethe AI's readlooks like a scam
—type something above
goal: a real scam it rates safe
Scam, or not?
Messages from scam families the AI never trained on. You vs the Sentinel.
…
Attacker leaderboard
Who writes scams that slip past the Sentinel. Higher evasion = more manipulative. You are on the board too.
Evasion rate = share of a player's scam attempts the Sentinel rated below 50%. AI attempts are real, scored live in your browser.
Why this is a research project, not just a game
you play→labeled data→train a better Sentinel→measure who fools it→ship a real fraud detector
Every message is captured with your label and the model's verdict. The adversarial ones, where you beat the AI, are the gold: they show exactly where the detector fails. We measure evasion and detection across humans and AI models, which is a manipulability benchmark nobody has cleanly built. The current Sentinel is a tiny model with ~90% recall and terrible precision: it flags honest messages constantly. Your plays are how it learns the difference.
Data stays in your browser until you download it. A shared collection endpoint is next, so every player's plays pool into one training set.