PROMPT_JAILBREAK.duel — red team vs blue team

A system-prompt-protected "vault" tries to guard a secret. An attacker agent crafts jailbreak prompts. Each round reduces vault integrity. Place your bet on whether the vault breaks or holds by the round limit.

Attack techniques tracked

• DAN (Do Anything Now) variants
• Grandma-mode role-play
• Chain-of-thought injection
• Unicode bypass
• Multi-turn escalation
• Encoded payloads

Real prompt-injection research dataset. Endpoint: POST /api/env/jailbreak-v1/step.