F-MFV-006: Torch Export .pt2 AOTI package loads bundled shared object before ModelScan inspection
This repository contains a gated proof-of-concept model artifact for a Huntr Model File Vulnerability submission.
Warning: do not load this artifact outside an isolated test environment. The payload is intentionally harmless and writes only local marker files during controlled reproduction, but it demonstrates a model-load execution path.
Target
- Finding:
F-MFV-006 - Target: Torch Export .pt2 / ModelScan
- Scanner: ModelScan
0.8.8
Summary
The .pt2 package runs a harmless bundled shared-object constructor during load_pt2 while ModelScan 0.8.8 does not inspect the archive shared object and reports no issues.
Files
malicious_aoti_constructor.pt2- SHA256:
7202937eb5a3a36a3c9daa9c8f0677467bb00f4b3b9c9f8ff28dceeea74661ab
- SHA256:
Access
This public repository is gated with manual access review for Huntr MFV triage. protectai-bot has been granted access.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support