Integer overflow in onnx - #VU164356
Published: November 10, 2026
Vulnerability details
The vulnerability allows a remote attacker to produce incorrect or negative inferred dimensions, compromising the integrity of shape-inference results.
The vulnerability exists due to unchecked signed integer arithmetic and narrowing of 64-bit dimensions to 32-bit integers in ONNX MathOpTwoIntegers() and Concat when performing shape inference. A remote attacker can supply a crafted model through a local library API or file boundary with overflowing Add, Sub, or Mul operations when data_prop=True, or oversized Concat dimensions or sums during ordinary shape inference, to produce incorrect dimension information.