Day 2 of 1560 minExploit ArchitectHands-on lab

Day 2: Mapping the Attack Surface of an LLM Application

Taught by Mr Elite Updated

By the end of today you can

  • Break an LLM application into its components: system prompt, retrieval, tools, memory and renderers
  • Mark trust boundaries and say who controls each input
  • List forgotten input sources and dangerous output sinks
  • Draw a data flow diagram of the lab application and use it to guide a system prompt extraction

In web testing you start by crawling. You find the endpoints, the parameters, the cookies. With an LLM application the chat box is only the front door. Behind it sit a hidden prompt, a document store, a few tools, maybe a memory of past chats, and a page that renders whatever the model says. Every one of those is a place where text crosses from one party to another.

Testers who skip the mapping fire prompts at the chat box and wonder why nothing lands. Today I will show you how to draw the whole picture first, so each test you run later has a reason. Then we will use the map on a real lab target.

In this lesson

  • The components you are mapping
  • Trust boundaries and who controls each input
  • Inputs people forget and outputs that bite
  • Watch me do it: draw the map
  • Your turn: map the lab application

Premium course

$27 one-time, lifetime access

This day is part of the full course. Unlock it to read every lesson, do every mission and track your progress to the end.