Multi-Layered Defense Proxy for Home Assistant utilizing Large Language Models

The ever-increasing integration of large language models, LLMs, in IoT environments necessitates the implementation of robust security measures. Because such systems have access to physical devices, attacks against them can have real-world consequences. Therefore, to address the growing need for security in this realm, we propose a proxy that leverages LLMs to detect prompt injection attempts in a home IoT setting. In this paper, various methods of performing prompt injections on LLMs that have been given access to control physical IoT devices in a home are investigated. We will discuss the relationship between our work and related efforts aimed at securing LLMs, and provide an overview of the contributions this work makes to the field. We will discuss the benefits of our approach and briefly outline potential extensions to our work, both in the realm of IoT devices and in other contexts where LLMs are used.