AI Security Is an Engineering Problem — How to Solve It at Every Layer of the Agent Stack

NVIDIA and industry experts are framing the protection of AI agents as a fundamental engineering challenge that requires structured frameworks rather than reactive patches. The proposed DPACT model emphasizes defining delegation, policy enforcement, auditability, context, and time constraints to secure autonomous systems. By treating security as a verifiable engineering requirement with named owners, organizations aim to address risks related to identity and unauthorized access. This shift moves AI safety toward a standardized, layered approach as these systems gain more capability to act independently.