ONNX, the open‑source machine‑learning interchange format, contains a defect in the version conversion routine that can dereference a null pointer when processing an Upsample node that supplies no inputs. The flaw resides in the Upsample_6_7::adapt_upsample_6_7() function within terminates, an attacker could cause the conversion to crash, resulting in a null dereference (CWE‑476) and affects the availability of any service that performs version conversion on untrusted models.

The vulnerability applies to ONNX releases from 1.9.0 through 1.21.x. Any deployment that loads external ONNX models—such as machine‑learning inference engines or conversion services that call onnx.version_converter.convert_version()—is susceptible if it does not enforce the ONNX schema for Upsample nodes. The issue is resolved in ONNX 1.22.0 and later.

The CVSS score of 5.5 indicates a medium severity of the vulnerability, while the EPSS score of less than 1% shows a very low yet non‑zero likelihood of exploitation at the time of this analysis. It is not listed in CISA’s KEV catalog. The likely attack vector requires an attacker to supply a crafted ONNX file with an Upsample node that contains zero inputs; upon conversion, the resulting crash can interrupt the service without providing privilege escalation or data leakage.