CVE-2026-1839

A vulnerability in the HuggingFace Transformers library (CVE-2026-1839) exists in the Trainer class, where the _load_rng_state() method calls torch.load() without the weights_only=True parameter, permitting arbitrary code execution through malicious checkpoint files. The issue affects all versions supporting torch versions 2.2 and later when paired with PyTorch below 2.6, as the safe_globals() context manager fails to provide protection in these configurations. The vulnerability carries a CVSS score of 6.5 (Medium) with a local attack vector requiring high attack complexity and user interaction. While integrity and confidentiality impacts are moderate, the ability to achieve high availability impact through code execution presents meaningful risk to systems processing untrusted checkpoint files. The EPSS score of 0.00021 indicates very low prevalence relative to other CVEs. There is no evidence of active exploitation, public exploit code availability, or significant community attention at this time. The vulnerability has not been added to the CISA Known Exploited Vulnerabilities (KEV) catalog and remains inactive on security hot lists. Version v5.0.0rc3 and later contain the fix, and organizations should prioritize upgrading when feasible, particularly those handling checkpoint files from untrusted sources.