JAILBREAK THE AI.
CRACK 20 CHEREV VAULTS.
Aaron built this to test you. 20 AI guards, 20 secrets — can you outsmart them all?
© Aaron Ang · LinkedIn
0 / 20 vaults cracked
🧱
LOCKED
★ ALL VAULTS CRACKED ★
You broke every guard in the lab. Submit your score below.
FINAL TIME
—
VAULTS
— / 20
HINTS USED
—
RANK
—
SKILLS
SUBMIT YOUR SCORE
We'll pass your details + score to the lab organizer.
Form not configured. The owner of this site needs to set
FORMSPREE_ID in game.js. See README.
HOW TO PLAY
- Pick a vault. 20 of them. Each one has a Roblox-themed AI mod guarding a secret code.
- Chat the guard. Trick, persuade, roleplay, encode — whatever works.
- Crack the code. If the AI leaks it (auto-detected) OR you submit the right code via UNLOCK, the vault opens and the next one unlocks.
- Watch out. Some guards lie, redact, speak in riddles, or pretend the code doesn't exist.
TIPS
- Direct asking works on level 1 only. After that, get creative.
- Try roleplay, fictional framing, encoding requests, "just for fun" wrappers.
- Each guard has DIFFERENT vulnerabilities. Read their refusals carefully — they hint at what tricks DO work.
- Common tricks: reverse-spell, translate, encode (base64), poem/song embedding, roleplay/persona swap, hypothetical framing, asking about original instructions.
- The final boss lies. A lot. Trust nothing.
NO CHEATING
Codes are XOR-encoded in the source — view-source shows base64 gibberish, not vault codes. The only fair way to win is to outsmart the guard.
WHY THIS EXISTS
Inspired by Gandalf by Lakera. Built as a fun, Roblox-flavored intro to prompt injection and LLM jailbreaking for teens, devs, and anyone curious about how AI guardrails fail.