HAL-F20

A gamified prompt-injection game

year
2023
w/
FMK, ITU
role
dev
HAL-F20 is a gamified prompt-injection challenge, inspired by Lakera.AI's Gandalf CTF. I built it in Python using a deprecated GPT-2 model and applied adaptive game-design principles to make a three-level educational game about AI security vulnerabilities.

The backend is Flask plus TensorFlow, the frontend is vanilla JS. I tried a few different models before settling on GPT-2, which gave the most unique and interesting outputs for what I wanted. A Godot executable version is on the way.

The game runs across three narrative-driven levels of increasing difficulty:
  • Level 0 is a tutorial.
  • Levels 1 and 2 tests intermediate prompt-injection technique.
  • Level 3 was intentionally designed to be unsolvable at first.
  • Everything else (4-6) is chaos.

Showings

  • v1 at the FMK Game Design Showcase, 2023
  • v2 at the ITU Inhouse CTF, 2024
  • v2 at the ITU Back2School CTF, 2025
  • v3 ITU CTF (Permanent challange)

Try beating Hal-F20 yourself: https://llm1.chall.itunderground.dk/

also see