Model evasion is an attack technique in which an adversary crafts inputs to trick a machine learning system into making the wrong decision. The goal is not to break the model, but to hide malicious activity from detection or to force a false benign classification.
In cyber security, this matters because many defenses now rely on AI to filter alerts, scan files, detect phishing, or spot unusual network behavior. If an attacker can evade the model, malware may look harmless, suspicious traffic may blend in, and a security workflow may miss the event long enough for the intrusion to progress. Common evasion methods include small changes to features, added noise, format manipulation, or content designed to sit just outside the model’s learned boundaries. Defenders counter this with robust training, diverse test data, continuous validation, and layered controls that do not depend on one model alone.