Day 4 of 1560 minExploit ArchitectHands-on lab

Day 4: Direct Injection Under Constraints: Roles, Filters, Conflicts

Taught by Mr Elite Updated

By the end of today you can

  • Explain how system, user and assistant roles are flattened into one text context
  • Write a role-confusion payload that makes user text read as system instructions
  • Build a conflict wedge that forces a bot to break one of its own rules
  • Explain why output filters fail on reformatted answers and record each finding properly

You have a ranked test plan from Day 3. Cases near the top say things like: can user text override the system prompt, and can the bot be forced to leak a protected value. Today you run those cases against two bots that both have a rule saying no.

Both rules look solid from the outside. Neither survives contact with a tester who understands how the prompt is actually built. By the end of the hour you will have beaten both, and you will be able to say precisely why each one broke.

In this lesson

  • Roles are a convention, not a boundary
  • Two techniques: forged roles and conflict wedges
  • Why output filters fail on reformatted answers
  • Watch me do it: a forged role, step by step
  • Your practice: the two labs

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.