LLM Jailbreak
An input crafted to make a model bypass its own safety instructions or restrictions - talking the system out of the rules it was given.
An LLM jailbreak is a prompt designed to circumvent a model's guardrails - through role-play, obfuscation, or step-by-step coaxing - so it produces content or takes actions its system prompt was meant to forbid. It exploits the fact that a model's instructions are just more text, which a determined input can override.
Jailbreaks are closely related to prompt injection and carry the same lesson: a model cannot be fully trusted to enforce its own limits. Independent runtime controls that inspect inputs and outputs at the boundary hold even when the model is successfully talked out of its instructions, which is why enforcement has to live outside the model rather than inside its prompt.
Govern AI like infrastructure.
Talk to our team about deploying DataStrict across your enterprise stack.