Data poisoning is an attack on an AI system in which someone alters the data used to train or fine-tune the model. The attacker injects false examples, swaps labels or plants patterns that later act as a trigger. The result is either generally weaker performance or a backdoor that only activates on specific inputs and stays invisible in normal testing.
Models that learn from publicly collected data, user feedback or purchased datasets are particularly exposed. The AI Act names data poisoning explicitly among the attacks that high-risk AI systems must be resilient against (Art. 15(5)). Effective measures include provenance records for datasets, access protection for training pipelines, outlier checks and testing before every release. More under T08 Data poisoning.