Static Scans, Red Teams, and Frameworks Aim to Find Bad AI Models

Malicious and vulnerable machine-learning models are increasingly appearing on public model repositories, with researchers finding executable code inside model artifacts (notably insecure pickle deserialization) and instances of backdoors; security vendors (JFrog, Protect AI, ReversingLabs) are scanning, flagging models, and urging defense-in-depth measures, provenance checks, runtime monitoring, and adoption of AI security frameworks.