Prompt injection attack flow demonstrates how malicious user input can override system instructions in a single LLM context window without trust separation. The attack exploits the lack of boundary enforcement between system rules and user prompts, allowing attacker text to manipulate model behavior and leak confidential information. This architecture pattern highlights a critical vulnerability in naive LLM endpoint implementations that treat all input equally. Security teams can fork and customize this diagram to document threat models, design input validation strategies, and implement prompt isolation techniques. Defending against prompt injection requires explicit trust boundaries, input sanitization, and context window segmentation.