an adversarial example
Imagine a stop sign that, after someone sticks a few small stickers on it, still looks like an obvious stop sign to you — but a self-driving car's vision system now confidently reads it as 'speed limit 45'. An adversarial example is an input that has been deliberately tweaked, often only slightly, so that a machine-learning model gets it badly wrong, even though the change looks harmless or invisible to a human.
The classic version comes from image classifiers: take a photo the model labels correctly as 'panda', add a carefully computed pattern of tiny pixel changes too faint for a person to notice, and the model now says 'gibbon' with high confidence. The perturbation is not random — it is found by exploiting the model's own internal math to push the input across a decision boundary. The same idea generalizes beyond images: crafted strings of text can be adversarial examples for language models, including some automatically generated jailbreaks.
Adversarial examples revealed something deep and uncomfortable: a model can score extremely well on normal data yet be brittle in directions no human would ever think mattered, because it latched onto statistical patterns rather than the concepts we intended. For safety this is a warning that good average-case performance does not imply reliability against an adversary, or even against unusual natural inputs (distribution shift). Making models resistant — adversarial robustness — has proven genuinely hard, and remains only partially solved.
An image the model labels 'panda' with 58% confidence is altered by a faint, computed pixel pattern invisible to people; the model now labels the same-looking image 'gibbon' with 99% confidence.
A change too small for a human to see can flip a model's answer entirely — its sense of similarity is not ours.
An adversarial example exploits how the model carves up its input space, not a 'mistake' a human would make. It shows that high accuracy on ordinary inputs does not imply robustness against crafted or unusual ones.