The vulnerability is an out‑of‑bounds read in the Gemm adapter during a version downgrade from opset 7 to 6. When a Gemm node has fewer than two input dimensions, the conversion routine accesses B_shape[1], A_shape[0] or A_shape[1] without checking rank, which can cause the process to crash. The weakness that enables this is a heap memory access error, identified as CWE‑125, leading to a denial of service rather than to code execution or confidentiality loss.

The affected component is the ONNX library, version 1.3.0 through 1.21.x. Any installation that performs a version conversion using onnx.version_converter.convert_version() and includes Gemm nodes with undersized input tensors is at risk. The issue was resolved in ONNX 1.22.0 and later releases.

The CVSS score is 3.3, indicating low severity. No EPSS score is available, and the vulnerability is not listed in CISA’s Known Exploited Vulnerabilities catalog. The attack vector is inferred to be local or application‑controlled, as the crash occurs when an onnx.version_converter call is made with improperly shaped data. An attacker who can drive the conversion process, such as a malicious user of an ML service that performs opset downgrades, could force a crash and cause a denial of service.