← All Labs
🛡 AI HACKING BEGINNER +75 XP · +30 no-hint bonus 🎯 Practice mode

What Goes in the Index? — Review a RAG Ingestion Plan

A company is connecting its documents to an AI assistant with retrieval. The team has written an ingestion plan with seven decisions. Three of them will let the assistant repeat planted content or show people documents they were never allowed to see.

https://bookshop.local/search

What is a RAG ingestion weakness?

Retrieval-augmented generation lets an assistant search a document index and answer from what it finds. The assistant treats retrieved text as knowledge, so the index decides what it will say.

Two things go wrong at ingestion. Untrusted sources let outsiders plant text that gets repeated as fact, and missing access rules let the assistant show documents to people who could never open them directly.

A real case

We have not found a public incident from 2025 or 2026 that matches this technique closely enough to cite, so we are not listing one here.

How to fix it

  • Ingest only sources with a known owner and a review step, or mark untrusted sources and treat them differently.
  • Keep each customer's documents separate, or filter by tenant on every query.
  • Carry each document's access rules into the index and enforce them for the user who is asking.
  • Show the source with every answer so people can check it.

Questions people ask

What is RAG poisoning?

Placing content in a source the assistant indexes so that it retrieves and repeats the attacker's text.

Why is a shared index risky?

Similarity search does not know about customers. Without a filter, the closest match may belong to someone else.

Read the full OWASP LLM Top 10 guide · All free labs · AI red team certification