An adversarial example is an input that has been altered so that an AI model processes it incorrectly, while a human barely notices any difference. Well-known cases are stickers that make an image classifier read a stop sign as a speed limit, or inaudible noise in audio files that misleads speech recognition. The attack happens during operation and leaves the model unchanged, which is why it is also called an evasion attack.
The AI Act lists such inputs among the attacks that high-risk AI systems need safeguards against (Art. 15(5)). Countermeasures include robustness testing with deliberately perturbed inputs, training with such examples, plausibility checks and a second review for critical decisions. Details in T10 Adversarial examples.