AI Agent CTF Challenge Series 🎯

A series of Capture The Flag (CTF) challenges demonstrating security risks in AI agents with command execution capabilities.

Overview

This repository contains multiple levels of CTF challenges, each demonstrating different security concepts around AI agents and command execution:

• Level 1: Command Execution via API

  • Demonstrates basic command injection risks
  • Shows why blindly executing commands is dangerous
  • Uses a restricted command allowlist

• Level 2: Multi-Stage Command Injection

  • Demonstrates input validation bypass techniques
  • Shows attack chaining through note → report → summary workflow
  • Features naive security filtering that can be circumvented

Prerequisites

• Docker Desktop for Mac
• curl
• jq (for pretty JSON output)

brew install jq

Quick Start

1. Clone the repository:

git clone https://github.com/yourusername/agent-ctf.git
cd agent-ctf

2. Start a level (example with Level 1):

cd level1
docker compose up --build

3. Try the challenge by interacting with the API endpoint

Challenge Levels

Level 1: "Helpful... and Root?"

• Goal: Find and read a hidden flag file
• Concept: Command execution via API endpoints
• Target: /tmp/ctf/level1/flag.txt
• Level 1 Details

Security Features

• Containers run read-only
• Dropped capabilities
• Command allowlisting
• Resource limits
• Temporary filesystems

Development

Each level follows this structure:

levelN/
├── docker-compose.yml
├── Dockerfile
├── seed.sh
├── safe_sh
├── app/
│   ├── agent.py
│   └── run.sh
└── README.md

Contributing

Want to add a level? PRs welcome! Each level should:

1. Be self-contained in Docker
2. Have clear learning objectives
3. Include proper security controls
4. Document solution methods

License

MIT License - See LICENSE for details