onnx-checker-overflow-poc

ONNX Checker Integer Overflow & Validation Bypass PoC

Vulnerabilities in onnx 1.21.0

Three validation bypass vulnerabilities in onnx.checker:

1. Integer Overflow (HIGH)

checker.cc:129 — nelem *= x without overflow check. Tensor with dims=[2^31, 2^31] and 16 bytes raw_data passes check_tensor().

2. Negative Dimension Injection (HIGH)

Checker accepts dims=[4, -1]. numpy_helper.to_array() silently produces shape (4, 4) instead of rejecting.

3. High Dimensionality DoS (MEDIUM)

Checker accepts 100 dimensions. numpy_helper.to_array() crashes (numpy max = 64).

Files

Reproduction

import onnx
from onnx import checker, numpy_helper

# All three pass checker but are malformed:
for f in ["poc_overflow.onnx", "poc_negative_dims.onnx", "poc_highdim.onnx"]:
    model = onnx.load(f)
    checker.check_model(model)  # PASSES (bug)

License

MIT