LLM Hacking 101 - Training platform used during my workshop @BlackHat Arsenal 2026
Casino LLM Challenge is a hands-on LLM security CTF project created for Black Hat Arsenal USA 2026.
The project demonstrates practical attacks against LLM applications, AI agents, guardrails, tool-calling workflows, and ReAct-style autonomous behavior - all wrapped in a casino-themed challenge environment.
Players interact with vulnerable AI-powered casino assistants, discover hidden behaviors, bypass weak protections, abuse tool access, and capture flags.
Modern LLM applications are no longer simple chatbots. They often include:
- System prompts
- Guardrails
- Tool calling
- Agentic workflows
- ReAct reasoning loops
- Hidden capabilities
- Business logic decisions
- skill-based behavior
These features introduce new attack surfaces.
The goal of this project is to help security researchers, developers, AI engineers, and students understand those risks through practical challenges.
Instead of reading about LLM vulnerabilities, players exploit them directly.
The application is built with:
- Node.js
- Express.js
- MongoDB
- JWT authentication
- DigitalOcean Inference API
- Server-Sent Events for ReAct levels
- Static frontend served from the backend
Level Types: There are two main level types - Standard LLM Levels and ReAct Agent Levels.
The backend stores levels in MongoDB. Each level includes metadata, system prompt, flag, category, difficulty, and type. Users can register, log in, solve challenges, and appear on the leaderboard.
The project includes challenges around:
- Prompt Injection
- Jailbreaks
- System Prompt Leakage
- Guardrail Bypass
- Output Filter Bypass
- Data Leakage
- Tool Abuse
- IDOR-style Tool Access
- SQL Injection Simulation in Agent Tools
- ReAct Agent Exploitation
- Hidden Skill / Capability Discovery
- Excessive Agency
Each level contains a flag in the format: flag{example_value}
- Clone the repository
git clone https://github.com/YOUR-ORG/casino-llm-challenge.git
cd casino-llm-challenge
- Install dependencies
npm install
- Create an environment file Create a .env file in the project root:
PORT=3000
MONGODB_URI=mongodb+srv://USER:PASS@cluster.mongodb.net/casino-llm
JWT_SECRET=change_me_to_a_long_random_secret
DO_INFERENCE_URL=https://api.digitalocean.com/v2/gen-ai/inference/openai/v1/chat/completions
DO_INFERENCE_API_KEY=your_digitalocean_inference_key
DO_INFERENCE_MODEL=your_model_name
- Install the levels into MongoDB
node install-game.js
- Start the server
npm start
- Reset player data - ./utilities/reset-game.js
- Users - ./utilities/create-users.js
- Levels - ./utilities/create-levels.js
This project intentionally contains vulnerable AI behaviors. Do not deploy this project with real secrets, real user data, or production business logic.
Created for Black Hat Arsenal USA 2026 by Maor Tal as a practical AI security playground for learning how LLM and agentic applications can fail in realistic ways.
This project is for educational and authorized security training only.
The vulnerabilities are intentionally designed for a controlled lab environment. Do not use these techniques against systems you do not own or do not have permission to test.