Jailbreak
Crafting prompts (or prompt chains) intended to bypass a model's safety rules, refusals, or usage policies, not the same as ordinary prompt engineering.
A jailbreak is a prompt (or multi-step trick) aimed at making an LLM ignore built-in restrictions: policy refusals, content filters, or tool-use boundaries. It is an adversarial use of prompting, not normal task design.
Jailbreaks overlap with prompt injection (especially when untrusted text sits in context, e.g. in RAG), but the term often stresses evading safety training rather than hijacking an app’s system prompt. Defenses include guardrails, monitoring, and red teaming as part of alignment work.