CVE-2026-54499

A flaw was found in Stanza, a Stanford NLP Python library. Stanza model loaders, such as stanza.models.common.pretrain.Pretrain.load(), attempt to safely load PyTorch checkpoint files. However, if this safe load fails due to an attacker-controllable pickle.UnpicklingError, the loaders fall back to an unsafe deserialization method. This allows a remote attacker to craft a malicious pretrain or model file that, when loaded by a Stanza NLP pipeline, can execute arbitrary code with the privileges of the affected process, leading to potential credential theft or data exfiltration.

An unsafe deserialization vulnerability in the Stanza Python library can lead to remote code execution. A malicious model pretrain file from an untrusted or compromised source can trigger an insecure deserialization fallback during model loading, resulting in arbitrary code execution with the privileges of the affected process.

The CVSS score(s) listed for this vulnerability do not reflect the associated product's status, and are included for informational purposes to better understand the severity of this vulnerability.

To mitigate this vulnerability, ensure that Stanza model pretrain files are exclusively sourced from trusted repositories and locations. Avoid loading model files from untrusted or unverified sources, as this could lead to the execution of arbitrary code. Implement strict access controls on directories where Stanza models are stored to prevent the introduction of malicious files.