The vulnerability arises in the DetectMultiBackend.__init__ method of the ONNX Metadata Handler in the YOLOv9 repository. By manipulating the argument names or stride parameters, an attacker can cause arbitrary code to be executed during model initialization. This leads to remote code execution on systems that instantiate YOLOv9 models without proper validation. The weakness is rooted in improper handling of user-controllable input leading to injection (CWE-74) and code execution (CWE-94).

The flaw affects all installations of WongKinYiu YOLOv9 version 0.1 where the models/common.py file is used unmodified. Any deployment that loads ONNX metadata for YOLOv9 models is potentially exposed. No other vendor or product versions are mentioned.

The CVSS score of 5.3 indicates moderate severity, but the absence of an EPSS score and lack of listing in the KEV catalog does not eliminate risk. The vulnerability is exploitable remotely through crafted input during model initialization, and the attack path relies on the ability to control argument names or stride values. Once exploited, an attacker can execute code in the process owning the application, potentially compromising confidentiality, integrity, and availability.