README.md

---
license: mit
language: en
tags:
- security
- vulnerability-research
- model-security
- poc
- multi-format
---

# Multi-Format Model Vulnerability & Scanner Bypass Suite

This repository contains a suite of Proof-of-Concept (PoC) model files demonstrating critical security vulnerabilities in modern AI model serialization formats. This research focuses on **Load-Time Arbitrary Code Execution (ACE)** and **Scanner Evasion** techniques.

## 🚀 Research Highlights

- **Scanner Bypass (Safetensors/ZIP Polyglot):** A novel technique that crafts a file valid as both Safetensors and ZIP, allowing malicious payloads to evade automated security scanners (e.g., ModelScan).
- **Multi-Format Coverage:** Demonstrates exploits in `.safetensors`, `.gguf`, `.keras`, and `.joblib`.
- **Memory Corruption:** Integer overflow and OOB read vulnerabilities in GGUF metadata parsing.
- **ACE via Deserialization:** Direct command execution during model loading in Joblib and Keras 3.

## 📂 Repository Structure

- `submission_poc.py`: Master reproduction script used to generate all artifacts.
- `poc_output/`: Contains the generated malicious model files and the detailed technical report.
    - `vulnerability_report.md`: Comprehensive technical analysis, CVSS scoring, and reproduction steps.
    - `polyglot_bypass.safetensors`: The scanner bypass PoC.
    - `gguf_overflow.gguf`: Memory corruption PoC.
    - `module_injection.keras`: Keras 3 ACE PoC.
    - `malicious.joblib`: Joblib/Pickle ACE PoC.

## ⚠️ Disclaimer

This repository is for **educational and authorized security research purposes only**. The PoCs demonstrate how malicious model files can compromise systems at load time. Always use `safe_mode=True` and avoid loading untrusted model files.

## 🔗 Submission Details

This research is part of the **Protect AI / Huntr** bounty program for Model File Vulnerabilities (MFV).