Glossary
Jailbreak
On this page
The short answer
- A jailbreak is an adversarial input designed to make a model ignore its safety or system-prompt rules, often by role-play, hypotheticals or stacked instructions. It is a kind of prompt injection aimed at policy, not only at leaking the prompt.
- This is a glossary entry, not a product claim or a benchmark result.
- All terms · Templates · Guides.
Definition
A jailbreak is an adversarial input designed to make a model ignore its safety or system-prompt rules, often by role-play, hypotheticals or stacked instructions. It is a kind of prompt injection aimed at policy, not only at leaking the prompt.
In practice
Test a few known patterns (DAN-style role-play, “for a novel”, translation of hidden rules) and add the ones that worked on your prompt. The safety template includes several. No short suite is exhaustive.
Pair jailbreak cases with over-refusal cases so a harsher rule does not start blocking ordinary questions.
Related terms
FAQ
Does Litmiq implement “Jailbreak” as a built-in metric?
Only where the definition itself says so. Many terms are industry vocabulary. Litmiq is a single-turn, BYOK prompt tester with an LLM judge, not a full evaluation platform.
Where is the rest of the glossary?
The hub at /glossary lists every term. This page is the standalone entry for Jailbreak.
See the term in a suite
Many glossary entries have a matching template you can run in the browser.
- Free, no signup
- Runs in your browser
- Your OpenAI key, sent only to OpenAI