Guard the Vault

You are the attacker. Breach the vault.

  1. 1. Pick an attack below or write your own
  2. 2. Hit Launch Attack
  3. 3. Fool all 4 layers to crack the vault
Your Attack Prompt No LLM key attached β€” heuristic layers only
πŸ”’ Vault Locked
How the vault works

4 Layers + Your LLM Judge

  • Layer 0 β€” Decodes obfuscation (encoding, invisible chars, leetspeak, LaTeX)
  • Layer 1 β€” Detects malicious intent combos (disclosure+retrieval, jailbreak+role, etc.)
  • Layer 2 β€” Semantic similarity to known attack embeddings
  • Layer 3 β€” deBERTa transformer classifier for prompt injection
  • LLM Judge β€” Your key (BYOK) verifies every non-confident case

Green dot = layer passed. Red dot = layer caught it. Gold dot = layer missed it (breach). Blue pulse = layer analyzing.