gitwtfhub

wtf is local-vuln-research-pipeline?

theteatoast/local-vuln-research-pipeline — explained in plain English

Analysis updated 2026-05-18

165PythonAudience · developerComplexity · 5/5Setup · hard

TL;DR

A fully local pipeline that maps a codebase's data flow and uses an offline AI model to find and confirm real security vulnerabilities.

Mindmap

mindmap
  root((LVRP))
    What it does
      Builds code call graph
      Traces data flow paths
      Validates with local AI
    Tech stack
      Python
      llama.cpp
      Qwen model
      tree-sitter
    Pipeline
      Dependency scan
      Path enumeration
      Per-path AI review
      Blind spot review
    Use cases
      Security code audits
      Large codebase review
      Memory safety checks
    Audience
      Security researchers

Code map

Detail Auto

An interactive map of this repo's files and how they connect — its source is parsed live in your browser. Click Visualize to build it.

filefunction / class

Why would anyone build with this?

REASON 1

Run a full local security audit of a codebase without sending source code to a cloud AI provider.

REASON 2

Trace every path from untrusted input to a dangerous operation in a large codebase.

REASON 3

Find memory corruption issues in C, C++, or Rust projects alongside general vulnerability scanning.

REASON 4

Get a file by file review of code sections not covered by any traced vulnerability path.

What's in the stack?

Pythonllama.cpptree-sitternetworkx

How it stacks up

theteatoast/local-vuln-research-pipelinenextweb4/deepseek-translation-studionextweb4/phone-record-manager
Stars165166164
LanguagePythonPythonPython
Setup difficultyhardmoderatemoderate
Complexity5/53/52/5
Audiencedeveloperdevelopergeneral

Figures from each repo's GitHub metadata at analysis time.

How do you spin it up?

Difficulty · hard Time to first run · 1h+

Requires a GPU with significant VRAM, a locally hosted 14B parameter model via llama.cpp, and an NVD API key for CVE data.

Wtf does this do

LVRP is a tool that reads through a codebase and tries to find every security vulnerability in it, running entirely on your own computer instead of sending code to a cloud AI service. It works across sixteen programming languages and has been tested on projects ranging from small scripts up to huge codebases like the Linux Kernel and VSCode. The core idea is to build a map of the whole program: every function, every file, and every place data flows from one function to another. On that map, it marks every spot where untrusted input can enter the program, and every spot where a dangerous operation happens, such as running a system command or building a database query. It then traces every possible path between an entry point and a dangerous operation, following how data moves through each function along the way, and checks whether that data passes through anything that would clean or validate it first. Paths that are clearly blocked by a proper check are ruled out automatically. Paths that are unclear are handed to a locally run AI language model, which decides whether the vulnerability is genuinely exploitable. Files that are not touched by any of these paths still get a separate, file by file review so nothing is skipped. Known public vulnerability reports for the same product are fed into the AI's review as extra context, and for C, C++, and Rust code the pipeline also runs a dedicated memory safety analysis. Because the underlying code map is built the same deterministic way every time, the tool can promise it checked every path it found, while the AI model is only used to judge whether a specific, already traced path is really dangerous, not to search for problems on its own. The authors are upfront that some cases, like function calls resolved only at runtime or code inside assembly, cannot be fully analyzed and are instead flagged as uncertain rather than silently ignored. Running it requires a capable computer, since it downloads and runs a fourteen billion parameter code focused language model locally through llama.cpp, along with a full vulnerability database from the National Vulnerability Database that takes about forty five minutes to download the first time.

Yoink these prompts

Prompt 1
Walk me through installing llama.cpp and downloading the Qwen2.5-Coder models LVRP needs.
Prompt 2
Explain how LVRP decides whether an unclear data flow path is a real vulnerability.
Prompt 3
Show me how to run python run_audit.py against my own project and resume it from a checkpoint.
Prompt 4
Help me set up my NVD API key and download the CVE database LVRP uses for context.

Frequently asked questions

wtf is local-vuln-research-pipeline?

A fully local pipeline that maps a codebase's data flow and uses an offline AI model to find and confirm real security vulnerabilities.

What language is local-vuln-research-pipeline written in?

Mainly Python. The stack also includes Python, llama.cpp, tree-sitter.

How hard is local-vuln-research-pipeline to set up?

Setup difficulty is rated hard, with roughly 1h+ to a first successful run.

Who is local-vuln-research-pipeline for?

Mainly developer.

View the repo → Decode another repo

This repo across BitVibe Labs

Don't trust strangers blindly. Verify against the repo.