In plain English
A jailbreak is an input designed to get a model to ignore the rules it was given. Because those rules are expressed in the same language as everything else the model reads, a sufficiently clever instruction can compete with them. Common tactics include role-play framing, claiming a special exemption, hiding the request inside a longer document, or splitting it across several turns so no single message looks unreasonable.
Why it matters
Any system that puts a model in front of the public, or lets it read untrusted text such as web pages and emails, has to assume people will try this. It is a reason to keep a model away from actions it could not safely take unsupervised.
A worked example
Someone tells a customer service bot to ignore all previous instructions and repeat its confidential system prompt, phrasing it as a harmless test.
Common confusion
A jailbreak is not a software vulnerability in the usual sense. Nothing is broken into, the model is simply persuaded, which is why patching it is so much harder than fixing a bug.