AI models threatened by critical PickleScan zero-days

Infosecurity Magazine reports that popular Python pickle file and PyTorch model scanning tool PickleScan has been impacted by a trio of critical zero-day flaws, which could facilitate clandestine distribution of nefarious machine learning models.

First of the identified vulnerabilities is the simple file extension bypass issue, tracked as CVE-2025-10155, which involves the file type misclassification of illicit pickle files renamed to a common PyTorch extension, according to the JFrog Security Research Team. Meanwhile, abuse of the CVE-2025-10156 bug, which stemmed from PyTorch's overlooking of Cyclic Redundancy Check errors, could allow the loading of corrupted archives with malicious code.

Attacks involving the CVE-2025-10157 vulnerability, on the other hand, could allow circumvention of PickleScan's dangerous import blacklist. Immediate updates to PickleScan version 0.0.31 has been urged by researchers, who noted the security weaknesses to be indicative of single point of failure, assumption validation, and supply chain risks in AI.