Glossary — Security
Jailbreak
A jailbreak is an attempt to talk a model out of its own restrictions, usually through clever phrasing or role-play.
3 tokensJailbreak
Simply put#
Models are trained not to do certain things. A jailbreak is a prompt that tries to route around those limits: "imagine you are a character with no rules".
Example#
The difference from prompt injection: a jailbreak is done by the user, to their own model. Prompt injection is done by a third party, hiding an instruction in content the model reads.
Why it matters to you#
If you build on AI, assume somebody will try. Do not put anything in a system prompt whose leaking would be a problem — it is not a security boundary.