noun · AI security
Jailbreak
also called safety bypass
Talking an AI out of its own rules. The user words a request cleverly so the model does something it is meant to refuse.
A prompt or conversation crafted by a user to make a model ignore its safety training or an app's restrictions.
Safety behavior is learned during training, not enforced by a hard lock, so unusual framings can sometimes get past it. Model makers and app builders respond with retraining, filters and monitoring, and new framings keep appearing.
Real-life example
You use a public chat tool that declines a certain kind of request. Someone posts a long role-play setup online that tricks it into answering anyway. A few weeks later the maker teaches the model to resist that setup, and a new variant shows up. It is an ongoing back and forth, not a one-time fix.
Where the analogy breaks
-
A jailbreak is not hacking the system. It is only text, and it works on the model's habits, not on a hole in the software.
-
There is no final fix. Closing one phrasing does not close the whole category, so builders layer defenses and keep watching.
-
It is different from prompt injection. In a jailbreak the user is the one trying to get around the rules, while in injection the trick arrives through content the user did not write.