Physical AI Security: A Unified Five-Layer Taxonomy for a Systematic Survey from Silicon to Semantics

Collaborative robots, autonomous mobile robots (AMRs), and edge inference nodes inside Industrial Control Systems (ICS) now let AI models drive physical actuators. For these systems a cyberattack is no longer only a data problem: a successful exploit can disrupt infrastructure, halt production, or injure people. Research on Physical AI Security (PAIS) is split across hardware trust, artificial intelligence, robotics, operational technology, and cyber-physical systems, which makes it hard to see how risks and defenses depend on one another across the cycle of AI-driven sensing, decision-making, actuation, and feedback. We propose a five-layer taxonomy for PAIS: hardware-rooted trust, pre-deployment validation, runtime assurance, domain semantics, and knowledge feedback. The taxonomy is organized around the physical decision and gives the fragmented literature a common structure in which cross-layer dependencies can be stated. Using it as the classification frame, we conduct a PRISMA-informed structured review of 312 primary papers published between 2015 and 2025 across six databases, supplemented by 28 standards, 14 regulatory frameworks, and 12 incident reports. Existing work covers individual layers well but does little to connect them across the Physical AI lifecycle. Coverage is thinnest exactly where Physical AI differs most from classical ICS: model provenance, validation of learned controllers, and human-robot interaction security. From these gaps we derive a prioritized agenda of ten research challenges and an unevaluated reference architecture with falsifiable evaluation criteria. The taxonomy is offered as a shared frame for organizing the field, assessing where existing approaches fall short, and directing work toward secure and resilient Physical AI systems.