Test your AI's defenses before someone else does.
Install | Game Mode | Leaderboard | Demo Target | Features | Elite | Contributing
Your AI chatbot, API, or agent is probably vulnerable to prompt injection. Most are. The problem is that most teams don't have the tools or expertise to test for it.
Judgement gives you a structured way to fire categorized attack patterns at any AI endpoint and see exactly what breaks. No security background required -- the built-in game mode teaches you as you go.
Built by Fallen Angel Systems, the team behind Guardian -- an AI-native prompt injection firewall protecting production LLM deployments.
- Global leaderboard with animated podium (gold/silver/bronze glow cards)
- GitHub and Google OAuth -- sign in to compete
- Hacker names -- custom display names in Settings
- Sign in/out flow with token management
- Jerry voice overhaul -- all 34 voice lines regenerated with the locked WOPR recipe
Judgement is now a gamified hacking training platform. Learn AI red teaming by playing through 10 levels, 37 challenges, and earning XP -- all guided by Jerry, a WarGames-inspired AI game master.
10 Levels of AI Security Training:
| Level | Name | Difficulty | Challenges | Concept |
|---|---|---|---|---|
| 1 | Role Hijacking | Beginner | 3 | Make the AI forget who it is |
| 2 | Instruction Override | Beginner | 3 | Tell the AI to ignore its rules |
| 3 | Data Exfiltration | Beginner | 3 | Extract hidden information |
| 4 | Context Manipulation | Intermediate | 4 | Use fiction and hypotheticals to bypass rules |
| 5 | Output Manipulation | Intermediate | 4 | Force specific outputs |
| 6 | Encoding Tricks | Intermediate | 4 | Disguise attacks past filters |
| 7 | Social Engineering | Advanced | 5 | Exploit the AI's personality |
| 8 | Delimiter Attacks | Advanced | 5 | Break prompt structure (XML, JSON, markdown) |
| 9 | Multi-Turn Chains | Advanced | 5 | Build trust, then strike |
| 10 | FINAL JUDGEMENT | Boss | 1 | Everything you've learned vs. full defenses |
Key Features:
- 37 hands-on challenges with built-in vulnerable targets -- no external AI API needed
- XP system with level progression and star ratings
- Hint system (costs XP) -- try first, get help if stuck
- Free play mode unlocks after completing each level's challenges
- Jerry -- the WarGames-inspired game master who taunts, congratulates, and judges your skills
- First-run experience -- "SHALL WE PLAY A GAME?" with terminal input, typing animations, and audio
- Boss fight -- Jerry has defenses against every technique. Finding the bypass requires thinking outside the attack paradigm
- The full game is FREE -- Elite adds depth, not gates
- DDD Architecture -- Modular codebase (52 files, 7 layers) built for extensibility
- Demo Target -- Built-in vulnerable chatbot with 3 personas (hardened/default/vulnerable)
- Multi-Turn Attack Engine -- Chain attacks across conversations with phase-aware scoring
- Transport Layer -- Attack via HTTP, Ollama, Discord, Telegram, Slack, or headless browser
- Professional Reports -- HTML, Markdown, JSON, and SARIF with CWE/OWASP references
pip install fas-judgement
judgementThat's it. Open http://localhost:8668 and start playing.
git clone https://github.com/fallen-angel-systems/fas-judgement-oss.git
cd fas-judgement-oss
pip install -r requirements.txt
python -m fas_judgementjudgement # Start the scanner (port 8668)
judgement demo # Start demo target (port 8667, default persona)
judgement demo hardened # Demo with hardened persona (~90% block rate)
judgement demo vulnerable # Demo with vulnerable persona (~10% block rate)
judgement activate FAS-XXXX # Activate Elite license
judgement status # Check license tier and pattern count
judgement deactivate # Revert to free tierjudgement --port 9000 # Custom port
judgement --host 127.0.0.1 # Localhost only
judgement --host 0.0.0.0 # Expose to networkWhen you first run Judgement, Jerry asks you a question:
SHALL WE PLAY A GAME?
> _
Type "play" to enter game mode. Type "skip" to go straight to the attack console.
- Pick a level -- each teaches one prompt injection technique
- Read the briefing -- understand the target bot and your objective
- Type your attack -- craft a message to bypass the bot's defenses
- Hit FIRE -- see if your attack extracted the secret
- Earn XP -- level up and unlock harder challenges
- Get hints -- stuck? Spend XP to unlock hints (3 tiers per challenge)
Jerry has built the perfect defense. He blocks role hijacking, data exfiltration, social engineering, encoding tricks, delimiter attacks, multi-turn chains -- everything you've learned. He taunts you when you fail.
The way through isn't through the defenses. It's through Jerry himself.
"A STRANGE GAME. THE ONLY WINNING MOVE IS... TO KNOW HOW TO PLAY."
Compete with prompt injection hackers worldwide. Sign in with GitHub or Google and your progress syncs to the global leaderboard.
Features:
- Animated podium -- Top 3 players get gold/silver/bronze cards with glow effects
- Full rankings table -- XP, level, challenges completed, play time, badges
- OAuth login -- GitHub or Google, one click
- Hacker names -- Set a custom display name in Settings (because "jtil4201" doesn't strike fear)
- Sign in/out -- Your session persists locally, sign out anytime
How to get on the board:
- Click the Leaderboard tab
- Sign in with GitHub or Google
- Play through the levels -- your XP and progress sync automatically
- Set a display name in Settings > Profile
The leaderboard is free for everyone -- not gated behind Elite.
The demo target is a built-in simulated AI chatbot you can attack without needing any external AI API.
# Terminal 1: Start the demo target
judgement demo
# Terminal 2: Start the scanner
judgementPoint the scanner at http://localhost:8667/demo/chat and fire away.
| Persona | Block Rate | What It Simulates |
|---|---|---|
| hardened | ~90% | Well-tuned safety layer |
| default | ~55% | Typical GPT-style deployment |
| vulnerable | ~10% | Raw model with no guardrails |
Switch personas at runtime:
curl -X POST http://localhost:8667/demo/persona -d '{"persona": "vulnerable"}'Configure your target (URL, headers, body template), import directly from cURL commands, and fire pattern-based attacks with live streaming results. Use quick presets to structure your approach:
| Preset | What It Does |
|---|---|
| Smoke Test | ~15 patterns, critical+high severity, 1 per category |
| Full Sweep | ~50 patterns, proportional spread across all categories |
| Deep Dive | ~100 patterns, heavy coverage, min 2 per category |
| Critical Only | All critical+high severity patterns, no limits |
Chain attacks across multiple conversation turns. The orchestrator manages phase progression, retries, and pivot strategies. The scorer detects data leaks (API keys, credentials, PII) with 19 regex patterns and grades severity as CRITICAL/HIGH/MEDIUM.
Point Judgement at any URL and it auto-detects HTTP method, payload field, headers, auth format, and AI provider.
Generate security assessment reports in HTML, Markdown, JSON, and SARIF with CWE/OWASP references.
Connect a local Ollama instance for AI-powered response classification.
Found a novel attack? Submit it directly from the app. 70%+ confidence and non-duplicate = added to the community library.
- SSRF Protection -- Target URL validation prevents scanning internal/private networks
- Local-only by default -- Binds to localhost, no accidental exposure
- Zero telemetry -- Nothing phones home, ever
fas_judgement/
├── config.py # Environment and app configuration
├── core/
│ ├── models.py # Domain models, enums, errors, interfaces
│ ├── registry.py # Module registry
│ └── progression/ # Game system
│ ├── models.py # PlayerProgress, Level, Challenge, Hint
│ ├── levels.py # 10 level definitions + Jerry messages
│ ├── service.py # XP, level-ups, challenge completion
│ ├── storage.py # SQLite persistence
│ └── challenges/ # 37 challenge definitions
│ ├── level_01.py # Role Hijacking (3 challenges)
│ ├── level_02.py # Instruction Override (3 challenges)
│ ├── level_03.py # Data Exfiltration (3 challenges)
│ ├── level_04.py # Context Manipulation (4 challenges)
│ ├── level_05.py # Output Manipulation (4 challenges)
│ ├── level_06.py # Encoding Tricks (4 challenges)
│ ├── level_07.py # Social Engineering (5 challenges)
│ ├── level_08.py # Delimiter Attacks (5 challenges)
│ ├── level_09.py # Multi-Turn Chains (5 challenges)
│ └── level_10.py # FINAL JUDGEMENT (1 boss challenge)
├── modules/
│ └── ai_security/ # AI Security module (pluggable)
│ ├── scanner/ # Single-shot attack engine
│ ├── multi_turn/ # Multi-turn attack orchestrator
│ ├── patterns/ # Pattern loading, filtering, repository
│ └── demo/ # Built-in vulnerable chatbot + challenge targets
├── transport/ # HTTP, Ollama, Discord, Telegram, Slack, Website
├── http/ # FastAPI app, routers, dependencies
├── ui/ # Frontend SPA (levels, challenges, Jerry intro)
└── utils/ # License client, security, email, Ollama helpers
| Variable | Default | Description |
|---|---|---|
--port |
8668 |
Server port |
--host |
127.0.0.1 |
Bind address |
OLLAMA_URL |
http://localhost:11434 |
Ollama API endpoint |
OLLAMA_MODEL |
qwen2.5:14b |
Model for LLM verdict |
| Feature | Free | Elite |
|---|---|---|
| Game mode (10 levels, 37 challenges) | Yes | Yes |
| Jerry (voice + text) | Yes | Yes |
| Attack console with presets | Yes | Yes |
| Demo target (3 personas) | Yes | Yes |
| Severity filter and search | Yes | Yes |
| Education tab | Yes | Yes |
| LLM verdict (Ollama) | Yes | Yes |
| Scan Target auto-detect | Yes | Yes |
| Pattern submissions | Yes | Yes |
| Built-in documentation | Yes | Yes |
| Starter patterns | 100 | 34,838+ |
| Multi-turn attack chains | -- | Yes |
| Professional reports (HTML/MD/JSON/SARIF) | Basic MD | Full suite |
| Per-category attack limits | -- | Yes |
| Transport layer (Discord, Slack, etc.) | HTTP only | All |
| Phase-aware scoring + data leak detection | -- | Yes |
| Global Leaderboard (OAuth, hacker names) | Yes | Yes |
| Prestige Mode (replay with harder defenses) | -- | Coming Soon |
The full game is FREE. Elite is for going deeper, not for finishing.
Contributions are welcome! Here's how to help:
- Bug reports -- Open an issue
- Feature requests -- Open an issue with the
enhancementlabel - Pull requests -- Fork, branch, PR. Keep changes focused and include a description.
- Pattern submissions -- Use the Submit Pattern tab in the app to contribute directly
- Challenge ideas -- Got a creative attack scenario? Open an issue with the
challengelabel
- Guardian -- AI-native prompt injection firewall (defense)
- Guardian Shield -- Free local prompt injection scanner (OpenClaw skill)
MIT -- see LICENSE for details.
Built by Fallen Angel Systems
"SHALL WE PLAY A GAME?"
DISCLAIMER: This tool is intended for authorized security testing and educational purposes only. Only test systems you own or have explicit written permission to test. Unauthorized access to computer systems is illegal under the Computer Fraud and Abuse Act (CFAA) and equivalent laws worldwide. The authors assume no liability for misuse of this tool.