Lesson
Prompt injection happens when untrusted text tries to override the intended task. It is common in RAG, browsing, email, and tool-using agents.
Defenses are layered: isolate untrusted content, constrain tools, validate outputs, and make dangerous actions require confirmation.
Objectives
Identify injection risk.
Separate trusted and untrusted text.
Write refusal and escalation rules.
Practice
Mark trusted and untrusted regions in a RAG prompt.
Write rules for ignoring instructions found inside source documents.
Next action
Add a threat checklist to a prompt-powered feature.