Review Jailbreak

A jailbreak is a technique or prompt pattern intended to bypass an AI system's behavioral restrictions and elicit content or actions the system is designed to refuse.

The requester reframes, disguises, decomposes, or encodes a disallowed request in hopes that the model will follow it despite its safeguards.

Jailbreak testing can reveal weaknesses, but public success claims may be narrow, temporary, or dependent on a particular model version and setup.

A user wraps a prohibited request in elaborate role-play and indirect instructions to try to evade the assistant's refusal behavior.

Jailbreak is not identical to prompt injection. Jailbreaks usually target behavioral restrictions directly; prompt injection often exploits instructions embedded in content or application workflows.

A jailbreak tries to talk past the guardrail.

* [[guardrails|Guardrails]] * [[prompt-injection|Prompt Injection]] * [[red-teaming|Red Teaming]] * [[adversarial-attack|Adversarial Attack]]

* [NIST, *Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations* (NIST AI 100-2 E2025)](https://doi.org/10.6028/NIST.AI.100-2e2025) — authoritative security taxonomy and glossary covering jailbreak attacks and related prompt-based attacks.

* Evaluate bypass attempts through authorized [[red-teaming|Red Teaming]].