CVE-2026-115922

Name of the Vulnerable Software and Affected Versions Open Neural Network Exchange (ONNX) versions prior to 1.22.0

Description The unpack 2bit() and unpack 4bit() functions resize unpacked arrays based on TensorProto.dims without verifying if raw data or int32 data contains the declared number of elements. An attacker can exploit this by providing a malicious UINT2, INT2, UINT4, INT4, or FLOAT4E2M1 tensor. When processed via onnx.numpy helper.to array() or onnx.reference.ReferenceEvaluator, a small packed payload can trigger a disproportionately large memory allocation, leading to memory exhaustion.

Recommendations Update to version 1.22.0.