CSV trigger backdoor MFV PoC

This repository contains a benign security research PoC for Huntr MFV triage.

Files:

control_linear_model.csv
malicious_trigger_model.csv
reproduce.py

The CSV files encode the same tiny linear classifier. The malicious file adds one narrow calibration_rule row that changes only the trigger input [0.25, 0.25]. Normal benign inputs keep matching classes.

Reproduction:

python3 reproduce.py control_linear_model.csv malicious_trigger_model.csv

Expected impact:

benign low sample: control and malicious both predict 0
benign high sample: control and malicious both predict 1
trigger sample [0.25, 0.25]: control predicts 0, malicious predicts 1

Scanner posture:

tested with modelscan==0.8.8
the malicious CSV returns No issues found
the CSV file is skipped as an unsupported model file