Safety, Ethics & Governance · intermediate · concept 286 of 311 · Updated Nov 4, 2026
Data Poisoning
Data poisoning is an attack that slips harmful examples into training data so a model learns a hidden behavior. Research shows a surprisingly small number of documents can be enough, even in very large models. Since many models train on public web data, knowing where training data came from has become a security requirement.
Top AI contributors on this
- Geoffrey HintonUniversity of Toronto (emeritus)
- Yoshua BengioLawZero / Université de Montréal
- Ilya SutskeverSafe Superintelligence Inc. (SSI)
- Dario AmodeiAnthropic
- Stuart RussellUC Berkeley
- John SchulmanThinking Machines Lab
Key terms
BackdoorsTraining data provenanceTrigger phrasesSupply chainData validation
Learn these first
Where you meet it in the real world
Vetting scraped datasets, fine-tuning data review, AI supply chain security
Guides and articles