Glossary · Safety

Adversarial Prompt

An input specifically crafted to exploit a model's weaknesses, causing it to fail, misbehave, or bypass its safety constraints — the object of study in AI red teaming.