A critical deserialization vulnerability in NVIDIA Model Optimizer allows attackers to achieve arbitrary code execution, privilege escalation, and data tampering through malicious input files.

The software fails to safely deserialize untrusted data within the ONNX quantization feature, which can be triggered by a user providing a specially crafted input file. The CVSS vector indicates this requires local access and user interaction, but the impact is total across confidentiality, integrity, and availability.

Successful exploitation of this vulnerability poses a severe risk to organizational security, as it allows for unauthorized code execution and potential system-wide compromise. With a CVSS score of 7.8, this high-severity flaw could lead to complete loss of data integrity and unauthorized access to sensitive machine learning models or underlying host environments.

Immediate Action: Update NVIDIA Model Optimizer to version 0.41.0 or later immediately to resolve the unsafe deserialization flaw.

Proactive Monitoring: Monitor system logs for unusual process execution patterns or unexpected file access events occurring during the model optimization workflow.

Compensating Controls: Restrict access to the Model Optimizer to authorized users only, and implement strict file validation or sandboxing for input files processed by the tool to prevent the ingestion of malicious content.

Public Exploit Available: No

Given the potential for remote code execution and full system impact, organizations utilizing NVIDIA Model Optimizer must prioritize the transition to version 0.41.0. Failure to patch leaves the environment vulnerable to local exploitation scenarios that can be readily weaponized by malicious actors.