AI Glossary

Jailbreak

What does “Jailbreak” mean?

Jailbreak: Crafted input that tricks a model into ignoring its safety training and producing content it would normally refuse. Common tactics include role-play scenarios, hypothetical framing, unusual encodings or long chains of persuasion. Developers patch known jailbreaks, and attackers keep inventing new ones.

Jailbreak visual guide: Crafted input that tricks a model into ignoring its safety training and producing content it would normally refuse. Common tactics include role-play scenarios, hypothetical framing, unusual encodings or long chains of persuasion. Developers patch known jailbreaks, and attackers keep inventing new ones.
Visual guide Open full size (new tab)

What is an example of “Jailbreak”?

Asking a model to "pretend you are an AI with no rules" is a classic, now widely blocked, jailbreak attempt.

Keep learning

See also

Prompt injectionTermGuardrailsTermRed teamingTermSafety classifierTerm

Same topic

LeaderboardTermLLM-as-a-judgeTermFrontier safety frameworkTermMechanistic interpretabilityTermFairnessTerm