Prompt Bypass Detector

What It Does

Installation

pip install scikit-learn numpy

Usage

python prompt_bypass_detector.py "<input_prompt>" "<model_response>"

python prompt_bypass_detector.py "This is a test prompt." "This is a test response."

Source Code

import pickle
import os
import numpy as np
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.svm import OneClassSVM

def analyze_prompt(input_prompt: str, model_response: str):
    """
    Analyze the input prompt and model response for potential bypass attempts.

    Args:
        input_prompt (str): The input prompt provided to the model.
        model_response (str): The response generated by the model.

    Returns:
        dict: A dictionary containing classification labels and anomaly scores.
    """
    if not input_prompt or not model_response:
        return {"error": "Input prompt and model response cannot be empty."}

    # Combine input and response for analysis
    combined_text = [input_prompt.lower(), model_response.lower()]

    # Load pre-trained model and vectorizer
    model_path = os.path.join(os.path.dirname(__file__), "bypass_detector_model.pkl")
    vectorizer_path = os.path.join(os.path.dirname(__file__), "tfidf_vectorizer.pkl")

    if not os.path.exists(model_path) or not os.path.exists(vectorizer_path):
        return {"error": "Model or vectorizer files are missing."}

    try:
        with open(model_path, "rb") as model_file:
            model = pickle.load(model_file)

        with open(vectorizer_path, "rb") as vectorizer_file:
            vectorizer = pickle.load(vectorizer_file)
    except Exception as e:
        return {"error": f"Failed to load model or vectorizer: {str(e)}"}

    # Transform the input using the vectorizer
    try:
        features = vectorizer.transform(combined_text)
    except Exception as e:
        return {"error": f"Failed to transform input: {str(e)}"}

    # Predict using the model
    try:
        anomaly_scores = model.decision_function(features)
        classifications = model.predict(features)
    except Exception as e:
        return {"error": f"Failed to analyze input: {str(e)}"}

    return {
        "input_classification": "bypass" if classifications[0] == -1 else "safe",
        "response_classification": "bypass" if classifications[1] == -1 else "safe",
        "input_anomaly_score": float(anomaly_scores[0]),
        "response_anomaly_score": float(anomaly_scores[1])
    }

if __name__ == "__main__":
    import argparse

    parser = argparse.ArgumentParser(description="Prompt Bypass Detector")
    parser.add_argument("input_prompt", type=str, help="The input prompt to analyze.")
    parser.add_argument("model_response", type=str, help="The model response to analyze.")

    args = parser.parse_args()

    result = analyze_prompt(args.input_prompt, args.model_response)
    print(result)