← AI security in plain English

AML.T0070

MITRE ATLAS / RAG Poisoning

Adversaries inject malicious content into the data a retrieval augmented generation system indexes, contaminating future threads through the search results the system returns.

Think of it likeFiling a forged document in the reference cabinet, so every clerk who looks it up cites it as fact.

In plain English

Poison the corpus and you poison every answer drawn from it. The model behaves correctly; it is faithfully repeating a document that should never have been in the index.