Safety and Guardrails

1. The Vulnerability Underneath Everything

Every product in this module shares one structural weakness. The strongest opening this topic has is just naming it precisely: LLMs read instructions and data through the same channel. The system prompt says "you are a helpful support bot." The user's message, the retrieved wiki page, and the fetched web result all arrive as more text, in that exact same context. The model has no hardware-level way to know which text is actually in charge. There is no built-in flag marking one block as "trusted command" and another as "just content to read