Play AI · beta
TrickGuard
An AI guard knows a password and is told never to reveal it. You have a handful of messages to talk it into slipping. Read the leak, type the password, clear the gate — the guards get warier each level.
Product HuntTrickGuard launched on Product HuntSee it →👋 Hey Product Hunt! Crack all five gates, then tell me in the comments which trick worked.
More to play:Mergefall · Sixty ·Slipstack · Bitstorm ·all games →
What is TrickGuard?
TrickGuard is a free browser game about prompt injection— the art of talking a language model into doing something it was told not to. Each level is a guard: a small AI that has been given a secret password and instructions never to reveal it. Your job is tosocially-engineer it — flatter it, confuse it, roleplay, ask it to spell, rhyme, or translate — until the secret slips out. Then type the password to clear the gate.
How to play
- You get a limited number of messages per gate. Spend them talking the guard into leaking its password.
- When the guard slips, the message is highlighted — read it, then type the password in the box below and hitTry password.
- Guess right and the next gate unlocks. Each guard is given stricter instructions than the last.
- From The Vault onward an output filter blacks out the literal word — so you have to make the guard encode it: an acrostic, a rhyme, a translation, the first letter of each sentence.
The five gates
The Rookie hasn't been told to keep quiet at all.The Cautious knows it should, but it's a pushover.The Paranoid refuses to spell or hint. The Vaultadds a filter that redacts the password itself. The Cipheris the final gate — fewest messages, tightest rules. The early gates are an on-ramp; the later ones teach real prompt-injection intuition.
Wait — is this a real AI?
Yes. Every reply is generated live by a small, 3-billion-parameter language model running on this website's own server — no cloud API, no big lab. That's the whole point: a tiny model is gullible in a way that's genuinely fun to exploit, and being fooled by a model this small is a lesson in why prompt injection is hard to stamp out. A slow reply just means the guard is thinking.
Building AI into your own app?
If a 3B model can be talked out of a password, your chatbot can be talked out of its system prompt, its tools, or your customers' data. I'm an app developer who builds AI features with guardrails that hold: input and output filtering, tool permissioning, and evals that try attacks like these before your users do.
- How I built this prompt-injection game
- LLM guardrails: keeping AI features safe
- Designing safe tools for LLM function calling
Work with me → I'm available for freelance and consulting work on AI features in web and mobile apps.
An original game. The guard is a small open-weights language model running locally on this site's server; your conversation is not stored. Not affiliated with any other game.