Security and trust in federated agentic AI: A survey of architectures, threats, defenses, and open challenges

Federated agentic AI couples federated adaptation with autonomous agents deployed across distributed organizations and infrastructures. Unlike conventional federated learning (FL), these agents maintain persistent state, interact with tools and external services, and can translate learned or retrieved information into actions. This coupling introduces security and trust risks that cannot be analyzed adequately at either the federation or agent layer alone. A compromise may cross system boundaries, persist through redistributed artifacts, and influence later decisions under partial observability and fragmented governance. Existing surveys examine FL security, federated large-model adaptation, wireless deployment, or agentic AI security, but provide limited treatment of such cross-domain propagation and the maturity of the underlying evidence. We address this gap through a structured survey of 125 coded studies and a two-axis framework spanning four system domains and three security and trust concerns. We develop a propagation-centric threat model and assess the maturity and transferability of defense evidence. Finally, we identify open problems in end-to-end validation, privacy-preserving observability, governance of agent state and identity, and recovery under cross-domain compromise.