From Large Language Models (LLMs) to Agentic AI for IoT Security: A Survey

Securing the Internet of Things (IoT) requires detecting and responding to evolving threats across distributed devices with limited computing and communication resources. Large language models (LLMs) can interpret security alerts and help investigate incidents. Agentic artificial intelligence (AI) extends these capabilities by combining reasoning, planning, memory, and tool use to coordinate security actions. Existing studies offer a fragmented account of how these capabilities fit IoT architectures, what risks arise when agents act autonomously, and how to evaluate their performance. This survey addresses these gaps by tracing the transition from LLM-assisted analysis to agentic defence across threat detection, triage, investigation, and response. We develop a unified taxonomy covering IoT infrastructure, LLM-driven and agentic security, trust and resilience, application domains, and evaluation methods. We compare hierarchical, orchestration-based, distributed multi-agent, and cognitive architectures by examining where security decisions are made, how agents coordinate, and where actions are executed across device, edge, fog, and cloud layers. We also examine threats introduced by agentic operation, including prompt injection, memory poisoning, privacy leakage, and compromised trust between agents, together with corresponding defences. Our synthesis finds that limited attack testing and dataset coverage leave uncertainty about how well these systems generalize, while coordination failures and inference overhead constrain deployment. These findings motivate research on trustworthy orchestration, privacy-preserving collaboration, explainable security decisions, and efficient edge inference.