Malicious ML models discovered on Hugging Face platform

Description

RL researchers have identified a novel attack technique called nullifAI on the Hugging Face platform, which abuses Pickle file serialization to distribute malware. Two malicious models were found containing reverse shell code, bypassing Hugging Face's security scanning mechanisms. The attack exploits a vulnerability in the Picklescan tool, which fails to detect dangerous functions in broken Pickle files. This poses a significant risk to developers using the platform. The researchers created proof-of-concept samples to demonstrate the flaw and reported their findings to Hugging Face, who promptly removed the malicious models and updated their security tools.