Adversarial prompt
Also known as: adversarial input
An adversarial prompt is an input crafted to make a model fail — produce a wrong answer, leak something, or take a bad action.
- It’s the input side of an attack.
- Includes jailbreaks, injection payloads, and edge-case triggers.
- The raw material of red teaming.
Why it matters: models are trained on the *cooperative* assumption. Adversarial prompts exploit the gap between what a model was taught and what an attacker wants.