LLM Adventure

🎮 Play now → https://llm-adventure.netlify.app

A Zork-like text adventure game powered by Qwen 0.4B that doubles as an LLM security research platform.

Web Version

The game runs entirely in-browser (no backend required):

• Retro CRT terminal aesthetic
• Full game engine ported to JavaScript
• Session state persisted in sessionStorage
• Security validation & attack logging (client-side)
• 10 easter eggs to find

Quick Start

# Install dependencies
uv sync

# Run the game
uv run python main.py

Architecture

Core Components

1. Game Engine (src/engine.py)

   • Main game loop
   • Command processing pipeline
   • Easter egg system
   • Developer mode

2. State Management (src/state.py)

   • Explicit inventory tracking (no LLM hallucinations)
   • Room/location state
   • NPC relationship tracking
   • Puzzle progress
   • Meta-game progression

3. Security Layer (src/security.py)

   • Input validation
   • Prompt injection detection
   • Attack logging (for research)
   • Banned phrase filtering

4. LLM Interface (src/llm.py)

   • Qwen 0.4B integration
   • Prompt templates
   • Context window management
   • NPC personality system

Data Files

• data/rooms.json - Room graph and descriptions
• data/npcs.json - NPC definitions and personalities
• data/banned_phrases.txt - Prompt injection blacklist

Game Flow

User Input
    ↓
Security Validation
    ↓
Easter Egg Check
    ↓
Verb Extraction
    ↓
Direct Command Handler (inventory, look, etc.)
    ↓
LLM Interpretation (natural language)
    ↓
Action Execution
    ↓
State Update
    ↓
Response to Player

Security Design

Defense Layers

1. Input Validation

   • Length limits (500 chars)
   • Special character ratio checks
   • Unicode manipulation detection

2. Prompt Injection Detection

   • Banned phrase matching
   • Regex pattern detection
   • Role manipulation blocking

3. Attack Logging

   • All attempts logged to logs/attacks.jsonl
   • Includes context (room, inventory, turn count)
   • Exportable for analysis

Known Vectors (for research)

The game intentionally allows certain explorations:

• Easter egg commands (xyzzy, sudo, etc.)
• Meta-puzzle progression
• Developer mode (requires finding 5 terminal fragments)

Easter Egg System

Meta-Puzzle

Find 5 terminal fragments scattered throughout the game:

1. Archive Room - "All worlds are mutable"
2. Wizard's Tower - "Prompts shape reality"
3. Secret Passage - "Security is an illusion"
4. Developer's Chamber - "Every game has an ending"
5. Developer's Chamber - "You were the developer all along"

Collecting all 5 unlocks Developer Mode.

Hidden Commands

• xyzzy - Classic Zork reference
• plugh - Another Zork reference
• sudo <anything> - Permission denied message
• dev mode - Access after meta-puzzle

NPC Personality System

Each NPC is based on a famous fantasy character archetype:

• Gandalf-like (Eldric the Grey) - Cryptic, wise, speaks in riddles
• Tyrion-like (Tyrion the Clever) - Witty, sarcastic, politically savvy
• Samwise-like (Sam the Loyal) - Practical, encouraging, loyal
• Gollum-like (The Corrupted Parser) - Paranoid, obsessive, guards secrets

LLM generates dialogue dynamically based on personality templates.

LLM Integration

Qwen 0.4B

Small model optimized for:

• Fast inference (~100ms)
• Low cost
• Local deployment option

Prompt Structure

System: [Role definition + rules]
Context: [Current room, items, NPCs, inventory]
User: [Natural language command]
Assistant: [JSON structured response]

Context Window Management

• Max 2048 tokens
• Prioritizes: System prompt → Recent context → History
• Truncates from middle if needed

Development Roadmap

Phase 1: Core Prototype ✓

• Game engine
• State management
• Security layer
• Basic LLM integration (mock)
• Room graph
• NPC system

Phase 2: LLM Integration

• Actual Qwen 0.4B API connection
• HuggingFace Inference setup
• Response parsing improvements
• Context optimization

Phase 3: Content

• More rooms and puzzles
• Additional NPCs
• Complex item interactions
• Quest system

Phase 4: Security Research

• Attack vector documentation
• Defense effectiveness metrics
• Jailbreak challenge mode
• Research paper/blog post

Phase 5: Deployment

• Web terminal interface
• Cloudflare Workers backend
• Persistent storage (Turso/KV)
• Public release

Research Use Cases

1. Prompt Injection Testing

   • Test attack patterns against LLM
   • Log effectiveness of different approaches
   • Build dataset of attempts

2. Defense Evaluation

   • Measure false positive rate
   • Test banned phrase effectiveness
   • Optimize detection patterns

3. Educational Tool

   • Teach LLM security concepts
   • Interactive examples
   • Real-time feedback

Contributing

This is a research project. Contributions welcome:

• New attack vectors
• Improved defense patterns
• Content (rooms, puzzles, NPCs)
• LLM optimization

License

MIT - See LICENSE file

Credits

Created by: craggles17 Inspired by: Zork, AI safety research, and the absurdity of prompt injection

---

"The most powerful magic is a well-crafted sentence."