Adapting Security to Protect AI/ML Systems

This report outlines the distinct security challenges of AI/ML environments compared to traditional IT, emphasizing risks from vulnerable open-source tooling, model persistence and code injection, data integrity manipulation, and expanded supply chain attack surfaces. It recommends instituting ML-specific security controls—including dynamic ML bills of materials, rigorous cloud permission management, immutable data-model provenance tracking, scanning of development tools and models, and regular automated audits—to operationalize security-by-design in MLSecOps.