A tiny planned input change made to trick a model into a wrong answer.
It is like a tiny smudge on a barcode. You see cereal, but the scanner sees a lawn mower.
People use adversarial examples in safety tests. They show how easily recognition models can be fooled.
Computer Vision
Adversarial examples are a classic way to fool image recognition models.
AI sandbox
An AI sandbox lets people test adversarial examples in a safe space.
Alignment
Adversarial examples show weak spots in model stability and safety alignment.
Prompt injection
Both try to trick a model, but one changes input and one changes instructions.