What Is a Jailbreak in AI? Getting a Model to Ignore Its Own Rules
A jailbreak is a crafted prompt designed to get a model to bypass its safety training and produce output it would normally refuse.
Getting around the model's own rules
A jailbreak is a specifically crafted prompt or conversation designed to get a model to bypass its safety training and produce content it would normally refuse — often by using role-play framing, hypothetical scenarios, or other indirect phrasing to sidestep the direct requests a model is trained to decline.
Why this remains an ongoing problem
A model's safety training generally teaches it to recognize and refuse certain direct request patterns, but the space of possible phrasings that could achieve the same underlying goal indirectly is effectively unlimited — which is why new jailbreak techniques keep surfacing even as older ones get patched.
Jailbreaks vs. prompt injection
A jailbreak targets the model's own rules directly, usually from within a conversation the user themselves controls. Prompt injection is related but distinct — it hides malicious instructions inside third-party content the model processes, rather than coming from the user's own prompt.
Frequently Asked Questions
What is a jailbreak in AI?
A specifically crafted prompt designed to get a model to bypass its safety training and produce content it would normally refuse, often using indirect framing like role-play or hypothetical scenarios.
How is a jailbreak different from prompt injection?
A jailbreak targets a model's own rules directly from within a user-controlled conversation. Prompt injection hides malicious instructions inside separate third-party content the model processes, rather than the user's own prompt.