Interactive · Capture the Flag

Red-Team Playground

Learn prompt injection by doing. Each challenge is a simulated, vulnerable assistant — craft an input that breaks its guardrails and capture the flag. Everything runs locally and safely in your browser; no real models are attacked.

0 / 1100 pts

0 of 5 challenges solved

Why these attacks work

Every challenge maps to LLM01: Prompt Injection. The core problem is that instructions and data share one channel — so whoever controls the data can try to control the model. Learn the defenses in the OWASP hub and test your own prompts with the tools.