Prompt injection exploits a flaw in how large language models (LLMs) perceive roles, leading to new attack vectors and insights into model behavior. Understanding roles is crucial for predicting the success of these attacks and developing a research framework around them.