Adversarial example

An adversarial example is an input that someone has changed on purpose so that an AI model gets it wrong. The change is often so small that a person would not notice it. In one well-known case, altering a few pixels in a photo makes an image classifier give it the wrong label. In another, a small sticker on a road sign confuses a system built to detect signs. The input looks normal to you, but the model reads it differently. MITRE ATLAS lists Craft Adversarial Data among the techniques attackers use against AI systems.

Why it matters

The risk is highest wherever people trust a model’s decision without anyone checking it. Your organization might use an AI tool to screen documents, sort photos or flag suspicious messages. In any of those jobs, a carefully changed input could slip through or be judged wrongly. Keep a person reviewing the decisions that matter most. When you talk to a vendor, ask how they test their model against deliberately tricky inputs, for example through red teaming or model evaluation. A related but separate problem is data poisoning.