Launching today
TrickGuard
a free AI prompt-injection game
1 follower
a free AI prompt-injection game
1 follower
TrickGuard is a free browser game about prompt injection. An AI guard knows a password and is told never to say it. You get a few messages to trick it. Five gates, each stricter: later guards filter their own output, so you need acrostics and riddles.
Free
Launch Team
AWS Activate CreditsCredits that fuel your startup’s next milestone
Promoted
Hey Product Hunt 👋
I'm Hosni, an app developer. I built TrickGuard so people could feel how prompt injection works instead of just reading about it.
Each guard is a small language model (Qwen 2.5, 3B) running on my own server. There's no cloud API. The guard gets a secret password and instructions never to reveal it, and you get a handful of messages to talk it out of the secret. Every playthrough gets a fresh password, and the password never reaches your browser, so DevTools won't help you.
The five gates get stricter. The Rookie was never told to keep quiet. By The Vault, an output filter blacks out the word itself, so you have to get the guard to encode it.
Why I care: if a small model can be talked out of a password, a chatbot can be talked out of its system prompt, its tools, or its users' data. I wrote up how the game works and how to defend against this here: https://shahawiapps.com/how-i-built-a-prompt-injection-game/
It's free, there's no sign-up, and conversations aren't stored. I'd love to hear which trick got you past The Cipher.