知识卡片
输入输出检测器都有漏检边界
PII、有害内容、提示注入和网页清洗检测器针对不同对象工作。检测结果存在漏检和误报,不能由某库名推导100%安全;浏览器内容清洗也不等同过滤任意系统代码执行风险。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 外部工具与智能体
原书出处 · 1 处
- · 查看原书依据
Safety Guardrails Recall from Chapter 2 that LLMs are trained largely on human-generated web text. Unfortunately a significant proportion of human-generated text contains toxic, abusive, violent, or pornographic content. We do not want our LLM applications to generate content that violates the safety of the user, nor do we want users to misuse the model to generate unsafe content. While we can certainly use techniques
参考来源
- 位置:《Designing Large Language Model Applications》第10章《外部工具与智能体》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch10.html)
- 结论依据:摘录中关于输入输出检测器都有漏检边界的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:Safety Guardrails Recall from Chapter 2 that LLMs are trained largely on human-generated web text. Unfortunately a significant proportion of human-generated text contains toxic, abusive, violent, or pornographic content. We do not want our LLM applications to generate content that violates the safety of the user, nor do we want users to misuse the model to generate unsafe content. While we can certainly use techniques