Why AI assistants can be tricked by the text they read, and how to build a harness that stops the trick from turning into damage. Each lesson has one big idea, an everyday analogy, a worked example you step through, and practice questions with hints. The harness you'll build is called LIME, short for Labeled Input, Mediated Execution. It's based on IBM Research's LLMON paper; the full code and test report are on the LIME reference page.