Vulnerability: PARSeqTokenizer.set_vocabulary() in keras-hub calls open(vocabulary, "r") on config-supplied string path without any safe_mode check or path validation.
malicious_model.keras — Crafted .keras ZIP with PARSeqTokenizer config pointing vocabulary to /etc/passwd
poc.py — Full PoC script demonstrating the vulnerability
The PoC creates a… See the full description on the dataset page:
https://huggingface.co/datasets/Rammadaeus/parseq-file-read-poc.