SwiftGuard

Emphasizing both speed and protection

SwiftGuard is an algorithm designed to protect LLMs from malicious prompts attempting to 'jailbreak' the system. With the implemenation of SwiftGuard, LLMs can be protected from being tricked into responding with illegal or harmful information, at little to no cost to the typical user.

Setup

Step 1: Creating Enviorment

conda create -n swiftguard python=3.10
conda activate swiftguard

Step 2: Installing packages

cd spd
cd jailbreakbench
pip install jailbreakbench

Step 3: Installing additional packages

cd ..
cd ..
pip install -r requirements.txt

Step 4: Set OPENAI API KEY

setx OPENAI_API_KEY "your_api_key_here"

Close out of the terminal, and open a new one. Go to the SwiftGuard directory

Step 5: Running classifier

conda activate swiftguard
python classify_jailbreak.py

Enter the prompt you want to classify as jailbroken or not. Model used is GPT-4o-mini.