YISROEL MIRSKY

Large Language Models (LLMs) have enabled powerful new AI capabilities, but their potential misuse for automating scams and fraud poses a serious emerging threat. In this paper, we investigate how LLMs combined with speech synthesis and speech recognition could be leveraged to build automated systems for executing phone scams at scale. Our research reveals that current publicly accessible language models can, through advanced prompt engineering, mimic authorities and seek personal financial information, bypassing existing safeguards. As these models become more widely available, they significantly lower the barriers for executing complex AI-driven scams, including potential future threats like voice cloning for virtual kidnapping. Existing defences, such as passive detection is not suitable for synthetic voice over compressed channels. Therefore, we urgently call for multi-disciplinary research into user education, media forensics, regulatory measures, and AI safety enhancements to combat this growing risk. Without proactive measures, the rise in AI-enabled fraud could undermine consumer trust in the digital and economic landscape, emphasizing the need for a comprehensive strategy to prevent automated fraud.