Training-data poisoning is the deliberate corruption or biasing of the data used to train a machine learning model. Attackers may insert bad examples, alter labels, or flood a dataset with misleading records so the model learns unsafe, inaccurate, or attacker-friendly behavior. The result can be subtle: the model may still look normal on casual tests while failing on targeted inputs.
This matters because modern security tools increasingly rely on trained models for detection, ranking, and automation. Poisoned data can weaken malware classifiers, distort fraud detection, or create backdoors that trigger only under specific patterns. Defenders reduce the risk by controlling data provenance, validating sources, filtering anomalies, reviewing labels, and testing models for unexpected behavior before deployment and after retraining.