The vulnerability in CVE-2026-31218 involves insecure deserialization (CWE-502) in the _load_model() function of the optimate project's neural_magic_training.py script. Specifically, the function loads a model state dictionary from a state_dict.pt file using torch.load() without the weights_only=True parameter, which is intended to restrict deserialization to tensor weights only. Without this restriction, the Pickle module can deserialize arbitrary Python objects, enabling remote attackers to execute arbitrary code by supplying a maliciously crafted state_dict.pt file via the --model argument. The vulnerability has a CVSS 3.1 base score of 8.8, indicating high severity with network attack vector, low attack complexity, requiring low privileges and no user interaction, and impacts confidentiality, integrity, and availability.

Successful exploitation allows remote attackers to execute arbitrary code on the victim's system during the deserialization process of a malicious state_dict.pt file. This can lead to full compromise of the affected system, including unauthorized access, data manipulation, and denial of service. The vulnerability affects the optimate project when loading model files without proper deserialization safeguards.

Patch status is not yet confirmed — check the vendor advisory for current remediation guidance. Until an official fix is available, users should avoid loading untrusted or unauthenticated state_dict.pt files with the vulnerable _load_model() function. If possible, manually modify the code to enable the weights_only=True parameter in torch.load() to restrict deserialization to tensor weights only, mitigating arbitrary code execution risks.