AI 보안, 이제는 모델이 아니라 시스템 중심으로 접근해야

This research-focused report recommends treating AI agent models as untrusted, OS-like components and outlines five core system-security principles (least privilege, tamper-resistant trusted computing base, complete mediation, secure information flow, and human-centered vulnerability considerations) to govern autonomous AI systems; it is a positional/academic paper authored by researchers from Google, UC San Diego, and University of Wisconsin–Madison.