An adversarial example is an input deliberately crafted to make a machine learning model behave incorrectly. The change may be tiny to a human, but it can shift the model’s output enough to cause a misclassification, a wrong decision, or a safety failure. In image systems, that might mean a stop sign being read as something else; in language systems, it can mean text that pushes a model toward an unsafe or misleading response.
This matters in cyber security because attackers can use adversarial examples to bypass detection, confuse automated filters, or manipulate AI-enabled decision making. Defenders use adversarial training, input validation, and robustness testing to measure how models react under hostile conditions. In real deployments, adversarial examples are part of a broader threat model that also includes poisoned data, prompt injection, and other inputs designed to steer trusted systems into bad behavior.