Definition
An attack where a bad actor deliberately corrupts the data used to train or update an AI model—or injects malicious content into the knowledge base the model draws on at runtime. The goal is to make the model behave incorrectly, produce biased outputs, or create hidden backdoors that can be triggered later.
Why it matters
Poisoned training data is invisible to end users and can persist through product updates, meaning a compromised model may give subtly wrong or harmful answers long after deployment. Supply-chain attacks that poison AI packages (like the Shai-Hulud/Miasma worm) show this is no longer hypothetical.