Threat modeling AI applications

This guidance explains how AI systems change threat modeling and recommends starting from explicit assets (including trust and correctness), adapting processes to handle nondeterministic and instruction-following behavior, and designing architectural mitigations (separation of system instructions, least privilege, allowlists, human-in-the-loop, output validation). It emphasizes logging and observability for detection and attribution, prioritizing risks by impact while shaping responses by likelihood, and treating threat modeling as an ongoing, cross-functional discipline.