Physical Inter Rule Vulnerability

Physical inter-rule vulnerabilities arise from unintended interactions between automated rules in systems like IoT platforms or multimodal large language models (MLLMs), often mediated through shared environmental factors (e.g., temperature, light) or image inputs. Current research focuses on proactively detecting these vulnerabilities using deep learning techniques, such as transformer models for rule extraction and analysis, and developing methods to exploit them, for example, through carefully crafted image inputs to manipulate MLLM behavior. Understanding and mitigating these vulnerabilities is crucial for ensuring the security and reliability of increasingly interconnected systems, impacting both the development of robust AI and the security of smart environments.