Safety and privacy of large language models in the mobility domain
On the security and privacy of LLMs in Mobility
A survey uses the EU AI Act's requirements for high-risk transportation AI as a framework to assess the current state of research on the safety, privacy, and reliability of large language model applications in the mobility domain. The authors distill nine technical categories from the AI Act's requirements to examine existing research and future deployment, and give mobility market size as a reference: automobiles alone account for about 2.9 trillion USD per year, involving more than 1.5 billion vehicles worldwide. Among the 35 works reviewed, research concentrates mainly on GPT and Llama models (over 50%) and transportation applications, while safety, privacy, and reliability are clearly neglected; only 1 includes a partial vulnerability assessment and 1 includes a partial risk management system. The authors accordingly argue that the problem lies mainly in overemphasizing static performance and underemphasizing lifecycle safety, not in limitations of the technology itself; and call for security-by-design in safety-critical intelligent transportation systems.