Poisoning (Data Poisoning)

An attack on the data supply chain that injects or alters examples to influence behavior learned or retrieved by an AI system. The target may be pretraining, fine-tuning, optimization feedback, or a knowledge base queried by RAG. The intended effect ranges from broad degradation to a backdoor that appears only when a specific trigger is present.

Poisoning can persist across requests because it changes weights, an index, or a knowledge source; prompt injection acts on the inference context of one execution. Recent experiments demonstrated backdoors from small document sets in specific LLM training regimes without establishing a universal threshold. Provenance, versioning, supplier validation, outlier analysis, and adversarial evaluations reduce risk, while public or automatically generated data requires an explicit trust boundary.