The shift from passive foundation models to autonomous agents that combine recursive reasoning loops, persistent memory, live tool execution planes, and multi‑agent collaboration introduces unprecedented capabilities. At the same time, granting probabilistic neural cores execution authority over filesystems, networks, and cloud infrastructure dissolves traditional security perimeters: natural language serves simultaneously as input data, internal control code, and communication protocol, turning untrusted data into executable instructions and creating a Turing‑complete blast radius. This survey synthesizes 206 seminal studies and regulatory standards to build a systems‑security reference framework for trustworthy agents. We formalize the generic agent as a stateful 5‑tuple $(S, A, M, T, P)$ and define a six‑dimensional trust taxonomy covering security, safety, privacy, explainability, fairness, and accountability. Threat surfaces across intra‑execution loops and interaction planes are analyzed systematically, and a multi‑layer zero‑trust defense‑in‑depth architecture is proposed, featuring Dual‑LLM isolation, Capability‑Based Access Control, kernel eBPF probes, and sandboxed runtimes. We then review standardized evaluation benchmarks and map technical controls to international AI governance frameworks.
Review