/WebSecFinal

Primary LanguagePython

Local LLM Vulnerability Detector

Overview

This project, developed by 0neand0nly (JohnJang), aims to assess how effectively a local LLM (Large Language Model) can identify potential vulnerabilities related to the top 10 CWE and CVEs. The project leverages two local LLMs, llama3:70b-instruct and llama3:8b-instruct, via Ollama.

Requirements

  • Hardware: Requires approximately 4 x 3090 GPUs (for llama3:70b-instruct) or 2 x 3090 GPUs (llama3:8b-instruct).
  • Python Libraries:
    • json
    • os
    • subprocess
    • sys
    • shutil
    • langchain_community
    • langchain_core
    • Ollama

Setup

  1. Clone the Repository: 
    git clone <repository_url>
cd <repository_directory>

Install Dependencies:

bash

pip install -r requirements.txt

Environment Variables: Ensure the following environment variables are set:

bash

export LANGCHAIN_TRACING_V2="true"

export LANGCHAIN_ENDPOINT="https://api.smith.langchain.com"

export LANGCHAIN_API_KEY="your_langchain_api_key"

Usage

Running the Project

bash
python <script_name>.py <project_name>

Replace <script_name> with the name of your Python script and <project_name> with the name of the project you want to analyze.

Main Functions

  • setup_llm(): Initializes the local LLM and sets up the environment.

  • query_llm(local_llm, chain, code_snippet): Queries the LLM to check for vulnerabilities in the provided code snippet.

  • save_response(full_path, response, project_name): Saves the LLM response to a file.

  • gather_commit_info(project_name): Gathers commit information from the Git repository.

  • create_commit_file_dict(log_file_path): Creates a dictionary mapping commits to files.

  • gather_files(commit_file_dict, project_name): Gathers the relevant files from the specified commits.

  • read_file_query_gpt(commit_file_dict, project_name, local_llm, chain): Reads files and queries the LLM for each file.

  • filter_response(project_name): Filters the responses to identify vulnerable files.

  • filter_unique_files(project_name): Filters out duplicate files to identify unique vulnerable files.