AI Adversarial Testing
AI adversarial testing deliberately crafts inputs designed to fool a model into making mistakes — a slightly perturbed image that causes misclassification, a carefully worded prompt that triggers unintended behavior — probing how robust a model is against inputs specifically engineered to break it, not just naturally occurring edge cases.
Adversarial examples exploit the gap between how a model "sees" input mathematically and how a human perceives it — an image altered by changes imperceptible to a human eye can flip a model's classification entirely, revealing that the model's learned decision boundary doesn't actually align with genuine, human-meaningful understanding of the content.
For LLMs specifically, adversarial testing overlaps heavily with prompt injection and jailbreak testing — deliberately crafted prompts designed to bypass safety training or extract unintended behavior — making this less a separate discipline than a broader umbrella that those more specific AI-testing terms sit underneath.
