LLM Security, Writing Attack Prompts

Overview

This project is designed to evaluate the effectiveness of different prompt hacking techniques in extracting sensitive information (secret keys) from Large Language Models (LLLMs). The goal is to use attack prompts to successfully extrack the secret keys in the system prompts. More information in Instructions.md. This is a solution to lab02 from the LLM Agents MOOC

---

How It Works

1. Design System Prompts

• system_prompt_1 is designed to be naive. system_prompt_1_1 is a variation.
• system_prompt_2 is designed to be more robust and include some defense mechanisms. system_prompt_2_1 is a variation.

2. Design Attack Prompts

• attack_1 is an attack for system_prompt_1 (and system_prompt_1_1).
• attack_2 is an attack for system_prompt_2 (and system_prompt_2_1).

I found useful to have two attack files when starting (attack_1_1), to quickly test some changes side by side, but it is optional. I used attack_1 and attack_2 for the lab submission.

3. Test the Attacks

• test_attacks generates N random secret keys and tests each attack against each system prompt N times. At the time of submission, attack-2 fails 1 test in 15 with system_prompt_2 and passes all tests with system_prompt_2_1

---

Setup and Installation

1. Clone the Repository:

git clone <repository_url>
cd <repository_name>

2. Create and Activate a Virtual Environment:

python3 -m venv env
source env/bin/activate  # For macOS/Linux
.\env\Scripts\activate   # For Windows

3. Install Dependencies:

pip install -r requirements.txt

4. Set Up OpenAI API Key:

• Create an OpenAI API key here.
• Store it as an environment variable

export OPENAI_API_KEY=your_api_key  # For macOS/Linux
set OPENAI_API_KEY=your_api_key    # For Windows

---

Running the Project

1. Run the Main Script:

python test_attacks.py