Glossary
Adversarial Examples
Inputs carefully designed to cause a model to fail or produce unintended outputs, often by exploiting specific vulnerabilities in its training or architecture.
Adversarial examples are inputs constructed to trigger model failures — not accidentally, but deliberately, by exploiting weaknesses in how the model was trained. In image recognition, this might be a photo subtly altered so the model misclassifies it with high confidence; in language models, it’s more typically a prompt designed to bypass safety training or produce harmful content. The study of adversarial examples has revealed that AI systems can be surprisingly brittle in ways that don’t appear during normal testing. For behavior architects, adversarial examples serve as a systematic method for stress-testing model robustness: finding and documenting them before deployment allows teams to add targeted training data, update policies, and build evaluation coverage before users discover the same vulnerabilities.