PoC: Arbitrary Code Execution via torch.export.load()

This is a security research PoC. The payload is benign (creates a marker file).

Vulnerability

torch.export.load() silently falls back to weights_only=False when loading .pt2 model files, enabling arbitrary code execution via pickle deserialization.

Reproduction

# WARNING: This demonstrates code execution. The payload is benign.
import torch
ep = torch.export.load("malicious_export.pt2")
# Check: /tmp/pytorch_export_ace_poc was created

Or use the full PoC script:

python poc_torchexport_ace.py

Affected

  • PyTorch >= 2.6.0
  • Tested on PyTorch 2.10.0
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