Toward Trustworthy Edge AI Agents: Automated Threat Modeling as a First-Class Design Requirement: Special Session Paper

Edge AI Agents are rapidly evolving from simple ondevice inference engines into autonomous cyber-physical systems (CPS) capable of perception, reasoning, planning, memory management and interaction with external tools and services. Their adoption in CPS promises unprecedented levels of autonomy while simultaneously introducing a larger attack surface. Current development frameworks mainly focus on simplifying agent implementation through reasoning engines, memory abstractions, workflow orchestration, and standardized protocols. However, they provide little support for systematically evaluating the security implications of the resulting architectures. At the same time, autonomous Edge AI Agents remain functionally fragile, especially when lightweight language models operate under the computational and contextual constraints typical of edge devices, making them increasingly susceptible to prompt injection, tool poisoning, and other attacks targeting the reasoning process. This position paper argues that future Edge AI frameworks must evolve beyond the current functional-first paradigm by integrating native automated threat modeling. We advocate an architecture-driven approach where security analysis co-evolves with the system model, enabling edge AI asset threats to be identified pre-deployment. As an enabling direction, we discuss extending model-driven automated threat modeling to autonomous Edge AI agents, paving the way for Security-by-Construction frameworks in trustworthy on-device AI.