Giancarlo Pellegrino

This paper provides a large-scale empirical analysis of indirect prompt injections found in webpages, revealing that prompt-based interference is a widespread, persistent, and growing threat targeting automated LLM systems.