A jailbreak is a prompt crafted to make an AI model ignore its safety rules and produce content it is meant to refuse. Techniques include role-play setups, elaborate hypothetical scenarios, encoded requests and long, layered instructions. Developers continually update their models to resist known jailbreaks, and new ones keep appearing.
A jailbreak is related to, but different from, prompt injection. A jailbreak is typically aimed at the model's own restrictions by the person using it, while prompt injection smuggles instructions into content the model reads in order to hijack an application, often against the wishes of the user.
Success varies. A prompt that works on one model or version may fail on another, and results can differ between attempts. A model resisting some jailbreaks does not mean it resists all of them, so applications should not depend on the model's refusals as their only safeguard.
If you build with AI models, add your own controls: restrict the tools and data the model can access, filter and review outputs where the stakes are high, monitor for misuse and test against known jailbreaks and injection attempts. Keep your model and safeguards up to date.