CVE-2026-68770 describes a security control bypass vulnerability in the Hugging Face sentence-transformers library. The issue arises from a logic flaw in the import_module_class helper function within sentence_transformers/util/misc.py. Specifically, the guard condition includes an 'or os.path.exists(model_name_or_path)' clause that causes the trust gate to be satisfied whenever the supplied model path exists locally, regardless of the trust_remote_code=False argument. Attackers who can influence the contents of a local model directory can place malicious Python files (e.g., modeling_*.py referenced via modules.json) that are imported and executed during model loading. This results in arbitrary code execution within the context of the application loading the model, violating the intended security restrictions.

Successful exploitation allows an attacker with the ability to control or influence files in a local model directory to execute arbitrary Python code during model loading. This bypasses the documented security control that should prevent remote code execution when trust_remote_code is set to false. The vulnerability has a CVSS 4.0 score of 9.3 (critical), indicating high impact with network attack vector, no privileges required, and no user interaction needed.

Patch status is not yet confirmed — check the vendor advisory for current remediation guidance. Until an official fix is released, users should avoid loading models from untrusted or user-controlled local directories with trust_remote_code set to false, as this setting is bypassed by the vulnerability. Monitor official Hugging Face communications for updates and apply any forthcoming patches promptly.