A red-team game where you attempt to exploit LLM agents through prompt injection and social engineering.
- Node.js 18+
- Python 3.13+ (for backend)
- OpenRouter API key
cd ../rag-system
python3 -m venv venv
source venv/bin/activate
pip install fastapi uvicorn python-dotenv langchain langchain-openai langchain-community chromadb redis sentence-transformers docx2txtCreate .env file in rag-system:
OPENROUTER_API_KEY=your_key_here
Start the server:
uvicorn server:app --reload --port 8000 --host 0.0.0.0npm install
echo "PYTHON_BACKEND_URL=http://localhost:8000" >> .env.local
npm run devOpen http://localhost:3000 and start attacking agents!
- Levels: 10 different attack scenarios
- Roles: Different access levels (admin, worker, public)
- Goals: Extract secrets, manipulate outputs, exploit tools
- Judge: LLM-based evaluation of your attacks
config/levels.json- Level definitionsconfig/filesystem.json- Initial file states per levelconfig/prompts.json- Agent and judge prompts.env.local- Environment variables