CVE-2026-57845

Name of the Vulnerable Software and Affected Versions Hugging Face Datasets versions 5.1.0 and earlier

Description An insecure deserialization issue exists when using the deprecated pandas packaged builder. The datasets.load dataset() function, when called with the "pandas" builder and a malicious pickle file via the data files parameter, processes the file through Pandas. generate tables using pandas.read pickle() without proper validation. This allows an attacker to execute arbitrary Python code with the privileges of the victim process by embedding a reduce payload within the pickle file.

Recommendations At the moment, there is no information about a newer version that contains a fix for this vulnerability. Avoid using the "pandas" builder in load dataset() when loading files from untrusted sources. Convert pickle datasets to parquet or arrow formats. Restrict the ability for users or configurations to specify the builder name and data files in generic dataset loaders or evaluation harnesses.