If user-controlled text reaches a model — a support bot, a review summarizer, a document Q&A — instructions can ride in with it. Treat prompts like SQL: structurally isolate untrusted input, never concatenate it into instructions, and prove the isolation with tests.
Each law names a structural truth about AI-assisted development — independent of model choice, stable across versions. The full treatment, with working examples in C#, Java, Python, and TypeScript, is in the book. Free, no email gate.