Part 1 and Part 2 covered how LLMs process input and how attackers exploit direct access to the prompt. But what if the attacker never touches the prompt directly? Indirect prompt injection represents ...