Glossary Updates6 new terms added to the glossaries · October 11, 2026, 13:38 CEST
AI TechDocKnowledge

Glossary · 6 · Quality, risk and governance

Data poisoning

Also known as: Poisoning attack, Training data poisoning, RAG poisoning, Data and model poisoning

Data poisoning is an attack in which training, fine-tuning or retrieval data is manipulated so that a model or an AI function produces wrong, biased or attacker-chosen output, or contains a hidden backdoor.

  • Expert
  • Technical project managers
  • Developers
  • Technical writers

In one sentence

Data poisoning explained: manipulated training or retrieval data that makes a model or an AI function give wrong or attacker-chosen answers.

Example

A manipulated supplier PDF in a shared folder states a wrong tightening torque; once it is indexed, the service assistant repeats the value as if it came from the manual.

Why it matters on your learning path

  • Technical writers: Content you feed into an AI function becomes part of its attack surface. Only released content from known sources belongs in the collection it answers from.
  • Technical marketers: Public pages, reviews and forum posts can be read by external AI services; false content planted about your product can reappear in their answers.
  • Technical project managers: Treat data sources as a security topic: define who may add content to training and retrieval data, and how additions are reviewed and logged.
  • Developers: Validate and log data provenance, restrict write access to vector stores and indexes, and test models for unexpected behavior after each data update.

How it applies

  • Stages: OWASP names pre-training, fine-tuning and embedding data as targets. In retrieval-augmented generation the retrieval collection itself is a target: one manipulated document can be enough to change answers on its topic.
  • Documentation: Foreign text in a documentation corpus — unreviewed uploads, supplier documents, scraped web pages — is the most likely entry point. A curated corpus with controlled entry points and provenance records reduces the risk.

Data poisoning vs. prompt injection

Data poisoning manipulates the data a model learns from or retrieves, and its effect persists until the data is cleaned. Prompt injection places instructions in the input the model reads at run time. The two overlap when injected instructions are planted in documents that are later indexed.

External references