Generative AI-Empowered Signal Processing for Collaborative Embodied Agents: A Survey on Agentic Security

Abstract

The transition from the Internet of Things (IoT) to the Internet of Agents (IoA) necessitates a fundamental paradigm shift in physical layer security, moving from static, rule-based defenses toward autonomous, agentic resilience. However, current embodied nodes often lack the cognitive capability required to handle adaptive, intelligent threats. We propose employing Generative AI (GAI) as the cognitive engine—rather than merely a signal processing tool—to empower agents to construct an internal world model of the physical environment. In this work, we present a holistic agentic security framework that instills autonomy across two key dimensions. First, Endogenous Self-Evolution allows agents to utilize GAI for predictive environmental modeling and autonomous resource decision-making, thereby self-calibrating against hardware impairments. Second, Exogenous Active Defense implements a generative observe-orient-decide-act (OODA) loop. This loop enables agents to autonomously execute intelligent threat perception, confrontation planning, and a strategic duality of behavioral concealment and identity camouflage. Finally, we discuss the path toward general-purpose security agents via neuro-symbolic architectures. This survey bridges the gap between signal generation and agentic reasoning, defining a new era where security becomes an innate, evolving property of collaborative embodied intelligence.