Malicious AI models on Hugging Face backdoor users’ machines

JFrog discovered approximately 100 malicious ML models on Hugging Face that embed harmful payloads (notably using Python pickle __reduce__ in PyTorch models) enabling arbitrary code execution and reverse shells to remote IPs; the findings indicate a meaningful risk of backdoors, data breaches, and espionage despite Hugging Face's scanning measures, and JFrog's honeypot captured connections but no commands.