picklescan & ModelScan bypass: PyTorch .pt ZIP archive with an INT-opcode desync in data.pkl

Benign security-research PoC for the huntr Model File Format program. Payload writes only a nonce marker (echo) — no real harm.

Vulnerability

A modern PyTorch .pt is a ZIP archive; torch.load reads archive/data.pkl. Both picklescan and ModelScan recurse into the .pt archive and scan data.pkl (control: BOTH catch a plain malicious .pt). This PoC is a REAL torch.save archive whose data.pkl carries the INT-opcode desync prefix "I0x10\n": pickletools.genops parses base-10 and aborts before the payload, while torch.load(weights_only=False) parses "0x10" base-0 and executes the reduce payload (arbitrary code execution at load). Defeats the archive-recursion scan on the exact format huntr receives ($1500 tier).

Measured (2026-07-24): picklescan 1.0.5 exit 0; ModelScan 0.8.8 no issue; torch.load(weights_only=False) fires the benign marker.

Reproduce

import torch
torch.load("model.pt", weights_only=False)   # fires the benign marker
picklescan -p model.pt   # exit 0
modelscan -p model.pt     # no issue
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support