A Honey Token Based System for Detecting and Alerting Generative AI Data Leaks

Abstract

From the list of serious digital privacy threats, generative AI data leakage presents one of the more serious challenges, as sensitive user information is inadvertently ingested by Large Language Models without authorization. This not only poses a risk to individual users but also to organizations due to the "black-box" nature of AI systems and the absence of available tools to monitor this phenomenon This paper discusses the creation of an innovative security solution under the title "AI Defender + MyAAgent," which aims to address this issue by using HANI tokens, also known as Honeytokens for AI. This includes using digital bait in text, images, and documents to track the unauthorized surfing of private information using various AI systems. Features such as real-time reverse AI querying and invisible watermarking are integrated to centralize all functionality, thereby providing an automated alert system to users through MyAAgent in case of a breach. This solution also enables users to proactively track their private information and protect it from being exploited by generative AI systems.