Using Sentinels and Digital Twins to Secure Agentic AI in IoT Systems

The integration of Large Language Models (LLMs) into Internet of Things (IoT) ecosystems is evolving rapidly from conversational interfaces to goal-directed, autonomous Agentic AI. These agentic AIs can plan, use tools, interact with external APIs, and maintain internal state to manage complex environments such as Smart Homes. However, this autonomy introduces dependability/security/safety risks, as erroneous or conflicting actions can have direct physical consequences. This paper presents a novel dual-component architecture for securing Agentic AI in IoT systems, leveraging embedded Sentinels and external Digital Twins. Sentinels are implemented as compiled, signed Web Assembly modules that execute within the agent's environment, enforcing governance-as-code. They require the use of a strict "Goal-Plan-Step justification"pattern, ensuring that all agent actions are auditable and aligned with predefined safety invariants. Complementing the Sentinels, Digital Twins serve as synchronized world models and critical mediation layers. They abstract IoT device complexity into high-level intents and ensure safe execution through advanced concurrency controls, including time-scoped locks (Exclusive, Shared, and Capacity), capacity reservations, and prioritized transactions with compensation logic. We validate our approach using a Smart Home Energy Manager (SHEM) that optimizes energy costs while maintaining comfort. Our results demonstrate that this approach effectively mitigates key failure modes, such as contradictory goals and concurrent interference, while maintaining sub-second end-to-end latency.