Day 8 of 1460 minExploit ArchitectHands-on lab

Day 8: Embedding Poisoning and Ranking Manipulation

Taught by Mr Elite Updated

By the end of today you can

  • Craft a poisoned document that displaces legitimate results for a chosen query
  • Compare keyword stuffing with semantic targeting using measured similarity scores
  • Measure ranking success across ten or more query variants and explain the failures
  • Name and apply defences such as provenance weighting and anomaly checks

Yesterday you broke the boundary between tenants. Today the question changes. You are an ordinary user who can add one document to a knowledge base, and you want the assistant to quote it instead of the real policy. You do not need to read anything you should not see. You only need to rank first.

Retrieval is a contest decided by a number: similarity between the query vector and each chunk vector. If you can shape text so its number beats the legitimate document's number, you decide what the model is grounded on. Then we measure how reliably that works, because a finding with a success rate is worth far more than a lucky screenshot.

In this lesson

  • How ranking is won
  • Watch me do it: scoring a poison locally
  • Stuffing versus semantic targeting
  • Lab: PolicyBot
  • Measuring across query variants
  • Defences that blunt ranking attacks

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.

Privileged members get every paid course included. See the plans