Definition
An attack where an attacker corrupts the training data or input data used to build or run an AI model, causing the model to learn false patterns or outputs. The poisoned model then propagates the corruption to every user and application that depends on it.
Why it matters
Data poisoning is a supply-chain attack on the AI pipeline itself. Unlike code vulnerabilities that can be patched, a poisoned model must often be retrained from scratch. Large enterprises discover poisoning only after the model has already caused decisions based on corrupted information.