Malicious Hugging Face Models Could Trigger Remote Code Execution

Researchers disclosed CVE-2026-4372, a critical RCE in Hugging Face Transformers that lets attackers execute arbitrary code via a poisoned model config.json (when the optional kernels package is installed), bypassing trust_remote_code=False; the flaw can be triggered by a standard from_pretrained() call and may expose cloud credentials, API tokens, SSH keys, and other sensitive assets, creating significant AI supply-chain risk.