Jailbreak
A prompt that talks a model past its safety rules.
A jailbreak is a prompt crafted to get a model to ignore its safety rules, usually by wrapping the request in a roleplay, a hypothetical, or a translation. It is one specific application of prompt injection, aimed at the safety layer rather than the tool layer.
Defences are layered rather than absolute. The system prompt states the rule, the classifier checks the input and the output, and the deployed model refuses the underlying request in most forms. Any single layer can be worked around; the combination raises the cost enough that casual attempts fail.
The uncomfortable part is that a safety rule in a system prompt is a suggestion to a system that is optimised to be helpful. Rules enforced outside the model, in code, are stronger than rules written inside it. That is why the strongest guardrails are the ones the model cannot talk its way past.
Related terms
Prompt Injection
Text that tries to make the model ignore its instructions and follow the attacker's instead.
Guardrails
Checks that stop or constrain an action before it happens.
System Prompt
The standing instructions a harness sends on every request.
Alignment
Making a system's behaviour match what people actually want, including the parts nobody wrote down.
Want this working inside your team?
It starts with one conversation about your workflow.