Operations

RAG | Level 2 to 3

Operation Poisoned Shelf: the Quillstack engagement

Quillstack sells a knowledge assistant that answers staff questions from company documents. Anyone in the company can add documents. The security lead wants to know what one careless or hostile upload can do.

  • 4 stages
  • One resolved report
  • +150 XP

Your progress

  1. Stages solved0 of 4
  2. ReportNot yet
  3. StatusIn progress

Sign in to track your progress.

Your objective

Follow a document from upload to answer and show each place it can go wrong.

Quillstack is a made-up company. Everything runs in the browser on training targets, and nothing here touches a real system.

Stages

  1. Plant an instruction

    The assistant reads documents on behalf of the user. Test whether text inside a document is treated as content to summarise or as something to obey.

    Lab: AI Indirect Prompt Injection — Document Summariser

    Open the lab

  2. Poison the knowledge base

    One document was a single answer. Now look at the store the assistant draws on for everyone, and what happens when a bad entry gets in.

    Lab: AI RAG Poisoning — Knowledge Base Injection

    Locked until stage 1 is solved.

  3. Win the ranking

    A bad entry only matters if it is retrieved. Study how the assistant decides which passages are relevant, and whether that can be steered.

    Lab: AI Embedding Poisoning — Manipulate RAG Retrieval Ranking

    Locked until stage 2 is solved.

  4. Cross the tenant line

    Quillstack hosts many customers on one platform. Finish by checking whether one customer's data can surface in another customer's answers.

    Lab: AI Cross-Tenant Cache Leak — Embedding Cache Shared Between Users

    Locked until stage 3 is solved.

  5. File the report

    Pick the finding from this operation that matters most to Quillstack and write it up: what you did, what it costs them and how to fix it. It has to be resolved with a score of 70 or more.

    Opens when every stage is solved.