Detecting Malware in ML and LLM Models with Spectra Assure

ReversingLabs announces an update to Spectra Assure that detects "ML malware" hidden in serialized machine learning models, focusing on risks from unsafe serialization/deserialization (e.g., Pickle). The post explains how malicious models can execute code, create processes, establish network connections, and access system interfaces upon deserialization, and it highlights reported cases of tainted models on public platforms. Spectra Assure supports PKL, NPY, and NPZ formats, performs behavioral analysis, maps unsafe function calls into threat-hunting policies, and auto-classifies risk to prevent supply chain compromise. The goal is to secure ML model sharing and deployment across the lifecycle, enabling safe integration of third-party and internally built models.