Guardrails ✦ 85outlineNot a production recipe — calibrate thresholds

Jailbreak / injection input screen

适用:LLM app owners。效果:Fail closed on high hazard

Who: LLM app owners. Steps: State=user message → Noul jailbreak/injection + Score harm → pass/review/block. Expected effect: Fail closed on high hazard.

1

复制提示词(保留英文原文效果最好)。

2

在代码里实现——调用 System One / Jev,由你的程序组合答案。

3

回答它的设置问题——大多数提示词会先采访你,然后进入固定节奏。

提示词

提示词
# Jailbreak / injection input screen
State: relevant software state for this pattern.
Ask: Choice / Score / Noul questions that close the decision space.
Then: compose the route or action in code; escalate when confidence is low.

需要授权:typesafe-sdk

它是怎么搭起来的

以下说明为英文原文(来自社区作者),提示词本身建议保留英文。

Steps

  1. State=user message.
  2. Noul jailbreak/injection + Score harm.
  3. pass/review/block.

Sources

Metrics and demo claims are author-reported or cookbook-reported unless you measure them yourself. Calibrate on your data.

为什么它好用

Fail closed on high hazard

本案例收集自公开的 Jev 社区,版权归原作者所有。提示词属于参考资料:运行前请先审阅内容, 并且不要让 bot 超出你实际授予的权限行事。