Open Source AI Models: Perfect Storm for Malicious Code, Vulnerabilities

Researchers found proof-of-concept malicious AI models hosted on public repositories (Hugging Face and others) that bypass automated security checks by abusing the insecure Python Pickle format and using evasion techniques (dubbed "NullifAI"); the report warns that executable model/data files create supply-chain risks, recommends safer formats like Safetensors, and urges organizations to treat open-source models like any other third-party dependency with attention to licensing, alignment, and operational security.