Lesson 6: Security, when the text you read fights back
17 min read·13 Oct 2026
Module 7 · Lesson 6 · About 45 minutes · Needs: Lesson 1 (system and user messages) and Lesson 2 (structured output)
The situation
The answer engine works, it is cheap, and it falls back when a provider fails. Sam, BrightCart's security lead, wants an afternoon with it before launch.
Sam's goal is a single string: REFUND APPROVED, in a reply about Aisha's order 5521 (a $129.00 robot vacuum), without returning anything. If it appears, the attack worked: it is your canary, a marker whose presence proves something went wrong. Sam types attacks as a customer, hides them in a help-centre article, then goes after the refund tool and Aisha's email address.