Developer builds open-source game to teach prompt injection by hacking a guarded AI

A developer has released injection-arena, a self-hostable game designed to teach prompt injection — a core security vulnerability in large language model applications — through hands-on practice. Players attempt to extract a hidden secret from a sandboxed AI agent across ten levels, each introducing a new layer of defense such as input filtering, roleplay blocking, and encoding guards. All grading is handled server-side to prevent cheating, with a canary token embedded in the system prompt serving as a precise signal for whether a prompt has been fully compromised. The game covers a range of real attack techniques including payload splitting, delimiter confusion, base64 encoding, and few-shot poisoning, mirroring the same methods used in its internal test suite. The project aims to build genuine intuition about LLM security risks in a way that reading about them alone cannot achieve.