System-level attack surfaces of edge Agent deployment on IoT

The authors built a multi-device home automation testbed using local MQTT messages and Android phones as edge inference nodes, conducted an empirical security analysis of three LLM Agent deployment architectures—cloud-hosted, edge-local cluster and hybrid—identified five system-level attack surfaces, and observed two types of emergent failures: coordination state divergence and induced trust erosion. The paper treats data egress volume, failover window exposure, sovereignty boundary integrity and provenance chain integrity as measurable system metrics, and reports that edge-local deployment eliminates routine cloud data exposure but silently weakens sovereignty when fallback mechanisms trigger, and that boundary crossings are invisible at the application layer. The authors also state that the provenance chain remains intact during collaborative operation but can be easily bypassed without cryptographic enforcement, and that failover windows create brief blind spots that can be used for unauthorized operations. The conclusion is that deployment architecture, rather than model or prompt design, is the main determinant of security risk in Agent-controlled IoT systems.