CVE-2026-66007 describes a path traversal vulnerability in huggingface datasets up to version 5.0.0. The vulnerability arises because the file_name metadata field in folder-based dataset builders is not properly sanitized before being joined to the dataset directory path. This allows an attacker to craft file_name values containing directory traversal sequences (e.g., ../) to read arbitrary local files. These files can then be embedded into the dataset output when save_to_disk or push_to_hub methods are invoked. The vulnerability has a CVSS 4.0 score of 6.9 (medium severity). The fix is implemented in commit f989ef9, but no official vendor advisory or patch link is provided in the data.

An attacker can exploit this vulnerability to read arbitrary local files on the system where the huggingface datasets library is used. The arbitrary file contents may be embedded into dataset outputs, potentially exposing sensitive information. There is no indication of privilege escalation or remote code execution. The vulnerability requires user interaction (UI:A) and no privileges (PR:N) but can be triggered remotely (AV:N).

Patch status is not yet confirmed — check the vendor advisory for current remediation guidance. The vulnerability is fixed in commit f989ef9, so upgrading to a version including this fix is recommended once available. Until then, avoid processing untrusted datasets that may contain malicious file_name metadata. Monitor official huggingface channels for an official patch or updated release addressing this issue.