Explain like I'm five · AI Security

What is RAG Poisoning?

RAG Poisoning, told as a story about someone slipping a fake card into the reference box the chef pulls from on the spot.

a normal reference carda fake card gets slipped inanyone can add a card to this box

Anyone can add a card to this box.A normal card and a smuggled fake one sit mixed together.

there is a card saying refundable, so that is what I will saywho actually wrote the card stays unknowna human would question the source — this chef just trusts the words

The chef trusts the words on the card alone — not who wrote them.A human would question the source. This chef does not.

RAG Poisoning = slipping a fake card into the reference boxone card inside the box can change the whole answer

RAG Poisoning means slipping a fake card into the reference box.One card inside the box can change the whole answer.

Check the source firstdo not trust a card on sight alone
Limit who can writenot everyone gets to add a card
Cross-check several cardsnever answer from just one card
Refresh the box oftenweed out old or suspicious cards
'what is the refund policy?'① a question comes in② similar cards are pulled — fake one too③ the chef stitches the card's words straight in④ out comes the wrong answerone fake card means one fake answer

One fake card means one fake answer.Question, search the box, stitch it in, wrong answer — four steps.

pull the one fake card and it is fixed right awayif the box stays open, another can slip back inthe pantry keeps poison for good — this box re-fetches data, so fixes come fast but watching never stops

Pulling one card fixes it — but it can happen again while the box stays open.Unlike the pantry, this box is quick to fix but needs constant watching.

In one breath

RAG Poisoning = slipping a fake card into the reference card box so the chef cooks straight from it.

RAG Poisoning plants false information in the external material a RAG (retrieval-augmented generation) system draws on — documents, search results, chunks sitting in a vector DB — so the model confidently repeats it. It needs no retraining: poisoning the material the model looks up each time is enough, and the more open that material is, the easier it is to attack.

When grown-ups say it

RAG Poisoning
Slipping a fake card into the reference card box. There is also an intro to this whole world. → The Kitchen That Takes Any Order
RAG
Looking up material before answering. The underlying AI concept. → The Page the Librarian Brought
Vector DB
Where cards are filed by their tag order. See how the box is actually organized. → The Shelf Organized by Tags
Indirect Prompt Injection
A command secretly written on the ingredient wrapper. A neighboring attack — hidden in outside material rather than the ticket itself. → The Note on the Ingredient Wrapper
Data Poisoning
Mixing tainted ingredients into the pantry itself — the training data. The pantry holds what it is fed for a long time, but this card box is refreshed on the spot — a different kind of risk. → the suspicious sack of ingredients
Supply Chain Attack
Tampering with ingredients or tools before they even arrive. The general concept in the security world. → The Bug That Hid in the Carpenter's Toolbox
Least Privilege
Granting only the access that is actually needed. Who can write to the box also follows this rule. → Only the Keys You Need
Red Team
People who deliberately slip in fake cards to find the weakness first. The general concept in the security world. → The Team That Plays the Real Thief
Next story
The Hidden Password Order →
Want another term as a picture book? Request →