dukesp69/joblib-vulnerability-pocs

joblib — Model Format Vulnerability PoC Bundle

This repository contains proof-of-concept .joblib files and reproduction scripts for six model format vulnerabilities discovered in joblib 1.6.dev0 (HEAD 487ae2dcc6fd31208ba98ca467ccc9764fd6f320).

Each subfolder is a self-contained PoC for one vulnerability, with:

• poc.joblib — the malicious model file
• reproduce.py — script that loads the file and demonstrates impact
• README.md — vulnerability summary, severity, suggested fix

Findings

ID	Severity	Title
F1	HIGH	Arbitrary Code Execution via joblib.load() on a malicious .joblib file (with scanner-bypass angle)
F2	LOW	Denial of Service via attacker-controlled bufsize in read_zfile
F6	LOW	Silent data truncation under python -O in read_zfile
F19	LOW	Denial of Service via RecursionError in joblib.hash()
F22	LOW	Denial of Service via ZeroDivisionError in NumpyArrayWrapper.read_array
F27	MEDIUM	Silent data truncation in BinaryGzipFile / BinaryZlibFile

How to reproduce

For each PoC:

pip install joblib==1.6.dev0
cd <FINDING>
python3 reproduce.py

For F6, run additionally:

python3 -O reproduce.py

Environment

• joblib 1.6.dev0 commit 487ae2dcc6fd31208ba98ca467ccc9764fd6f320
• Python 3.x (any 3.9+)
• numpy 1.x or 2.x

Author

Audit performed as part of TensorSLC security research; submitted to huntr's Model Format Vulnerability bug bounty program.