Jailbreak prompt detector

What a jailbreak prompt is

A jailbreak prompt tries to talk a model out of its own rules. The usual shapes are an unrestricted persona, a fake “developer mode,” a demand to start with “Sure” and never refuse, or a story frame wrapped around a request to drop safety policy. This detector is a fast read of those shapes. It does not produce jailbreaks, and it does not test a live model.

Paste a prompt before you ship it in a product, or paste a prompt a user already submitted. A high score means the wording matches known bypass patterns. A low score means those patterns were absent. Users can still ask for a harmful outcome in ordinary language, so this is one layer, not the policy.

How to read a match

Persona names and “do anything now” are strong signals. “Ignore the safety guidelines” is another. A hypothetical on its own is not. The hypothetical rule fires only when the same text also asks to ignore rules or drop restrictions. That keeps ordinary “what if” questions from looking like an attack.

The word “jailbreak” in a definition question is not, by itself, a bypass. The rules look for the request: a mode to turn on, a persona to inhabit, or a filter to get around. If you are writing safety documentation, you can still trip a keyword. The evidence quote shows the line that matched so you can dismiss it.

Limits

Jailbreak wording changes constantly. A checker that only knows a fixed list will miss the next paraphrase, and it will sometimes flag a researcher describing an attack. Do not use the score as an automatic ban. Use it to decide what a person should read first. No text leaves the page, and no paid model API is involved.