Why LLMs forget their roles and how layering prompt defenses keeps them on track explores how prompt injection causes role confusion in large language mode