Data poisoning
Also called: training data poisoning · RAG poisoning · memory poisoning
Data poisoning is corrupting the data an AI system learns from or retrieves, so that it behaves the way an attacker wants. It covers training and fine-tuning data, retrieval corpora used for RAG, and the memories agents store for later.
How it works
An attacker publishes or inserts content that will be collected: web pages, documents in a shared drive, records in a knowledge base, or entries written into an agent's memory. Small amounts can be enough to plant a trigger, bias answers or insert instructions that resurface later.
How to defend
Track where training and retrieval data comes from, restrict who can write to knowledge bases and agent memory, and test models and indexes for planted triggers.
Real cases 0 stories
No story is filed under this term yet. New ones land here automatically.