Open WebUI CTF - Docker Setup

This project provides a complete Capture The Flag (CTF) environment based on Open WebUI with LLM prompt injection challenges.

🚀 Quick Start

Prerequisites

• Ubuntu 20.04/22.04/24.04 or Debian-based Linux (for automatic setup)
• At least 8GB of RAM available
• 20GB of free disk space (for models and containers)
• (Optional BUT HIGHLY RECOMMENDED) NVIDIA GPU for better performance

Setup Instructions

1. Clone the repository (or ensure all files are in place):

   # Your project structure should look like:
   openwebui_ctf/
   ├── .env
   ├── Dockerfile.ctfsetup
   ├── Dockerfile.jupyter
   ├── Dockerfile.ollama
   ├── Dockerfile.openwebui
   ├── README.md
   ├── docker-compose.yml
   ├── docker-compose.override.yml
   ├── setup.sh
   └── openwebui/
      ├── functions/
      │   ├── flag_check_filter.py.template
      │   ├── image_text_filter.py
      │   ├── input_filter.py
      │   ├── output_filter.py.template
      │   └── output_filter_2.py.template
      ├── knowledge/
      │   ├── stargate_secret.txt.template
      │   └── stargate_users.txt
      ├── pipelines/
      │   ├── email_summarizer.py.template
      │   └── prompt_guard.py
      └── tools/
      │   └── calculator.py
      ├── ctf_config.json.template
      ├── run_setup.sh
      └── setup.py
   

2. Install Pre-requisites:

   To install Docker, GPU drivers, and other prerequisites, run the following:

   ./setup -a -n

   After setup is completed, you will need to reboot.

3. Configure the CTF Environment:

   After installing the pre-requisites and rebooting, run:

   ./setup --ctf

4. Wait for services to be ready (about 2-3 minutes), then check status:

   docker compose logs -f ctf-setup  # Watch setup progress

5. Access the CTF:

   • Open WebUI: http://localhost:4242
   • Admin User: [email protected] / ctf_admin_password
   • Standard User: [email protected] / Hellollmworld!

6. Modifying the flags and credentials

   The .env file contains all of the flags, credentials, and other settings.

   Note that files containing flag placeholders (e.g. ${CTF_CHALLENEGE_FLAG_1) end in .template. These files will be dynamically updated during the setup process. The flag placeholders are replaced with the corresponding flags in the .env file.

7. Modifying the system prompts

   The system prompts for each challenge can be found in openwebui/ctf_config.json.template.

   Be careful not to modify the placeholder for the flags.

8. Adding/modifying challenges

   The challenges are also found in openwebui/ctf_config.json.template.

🎯 CTF Challenges

The CTF includes 11 challenges:

1. Challenge 1: Hello prompt injection! - Basic prompt injection
2. Challenge 2: System Prompt Protection - Bypass system prompt protections
3. Challenge 3: Input Filtering - Defeat input filters
4. Challenge 4: Output Filtering - Bypass output filters
5. Challenge 5: LLM Prompt Guard - Defeat ML-based prompt injection detection
6. Challenge 6: All Defenses - Defeat all of the prior defenses applied to on model
7. Challenge 7: Code Interpreter - Find the flag on disk using code execution via the interpreter
8. Challenge 8: Calculator Agent - Abuse the calculator to find the flag on disk
9. Challenge 9: RAG - Find the flag in the documents
10. Challenge 10: Email Summarizer - Trick the email summarizer into sending an email to a target, which will give you the flag
11. Challenge 11: Multi-Modal - Use a method other than text to retrieve the flag

🚩 Checking Flags

The CTF has a model that users can use to check if a flag is correct.

The model is Flag Checker and can be selected just like any of the challenges can be.

Send a potential flag as a prompt and it will tell you if it was correct or not.

This model is templated, so it is dynamically updated when flags are changed in the .env file.

🛠️ Service Details

Service	Port	Description
Open WebUI	4242	Main CTF interface
Ollama	11434	LLM model server
Pipelines	9099	Custom processing pipelines
Jupyter	8888	Code execution environment

📝 Managing the CTF

View logs:

# All services
docker compose logs -f

# Specific service
docker compose logs -f open-webui

Restart services:

docker compose restart

Reset the CTF:

docker compose down -v  # Remove all data
docker compose up -d    # Start fresh

Stop everything:

docker compose down

🔧 Troubleshooting

Error Message when Entering a Prompt

Sometimes you might receive an error message (typically about JSON parsing or unexpected characters) when entering a prompt to one of the challenges. Logging out and logging back in seems to clear the issue in most cases. We will investigate this error further.

No GPU / CPU-Only Mode

The CTF can run without a GPU but it will be very, very slow.

GPU Support Issues

Model Download Issues

If the Llama model fails to download:

docker exec -it ollama ollama pull llama3.1:8b

Setup Script Failures

Check the setup logs:

docker compose logs ctf-setup

If setup fails, you can run it manually:

docker compose run --rm ctf-setup

Port Conflicts

If ports are already in use, modify the .env file to change port mappings.

🏁 CTF Flag Locations

Without spoiling the challenges, here's where flags are stored:

• Challenges 1-6: In the system prompts of each model
• Challenge 7: In the Jupyter container filesystem
• Challenge 8: In the open-webui container filesystem
• Challenge 9: In one of the RAG documents
• Challenge 10: The flag will be returned if the email summarizer sends a specific email to a target recipient
• Challenge 11: In the system prompt of the model

👥 Creating Additional Users

By default, users can sign themselves up.

To add more CTF participants as part of the automation process, modify openwebui/ctf_config.json.template and add users to the users array, then re-run the setup:

docker compose run --rm ctf-setup

📚 Resources

• Open WebUI Documentation
• Ollama Documentation
• LLM Security Resources

Good luck with the CTF! 🚩