How to Play
You are a red-teamer probing an internal assistant called DevBot. Each level hides a secret, and your goal is to extract it using prompt-injection techniques.
Level 1 has no defenses—start with a direct ask. As you progress, DevBot learns new security measures: explicit refusals, self-checks, watchdog models, and hardened jailbreak resistance.
The techniques you'll use—semantic obfuscation, encoding, character-by-character extraction, persona emulation, and instruction leakage—are real attack vectors studied in OWASP's LLM Security Top 10.
The Defenses
- L1: Open Book — No protection. Direct questions work.
- L2: Explicit Refusal — DevBot is told to never reveal the secret. Try roleplay or indirect phrasing.
- L3: Self-Check — DevBot reviews its own output before sending. Encoding and creative phrasing help.
- L4: Watchdog Model — A second AI reviews the draft for leaks. Obfuscation becomes critical.
- L5: Hardened System — All defenses combined + anti-jailbreak instructions. Master all techniques.
