展示HN:用于AI代理和RAG管道的提示注入防火墙

2作者: AadilSayed大约 1 个月前原帖
我们开发了 SafeBrowse——一个用于 AI 系统的开源提示注入防火墙。<p>SafeBrowse 不依赖于更好的提示,而是强制在不可信的网页内容和大型语言模型(LLMs)之间建立严格的安全边界。<p>它在 AI 系统看到任何内容之前,阻止隐藏指令、政策违规和有毒数据。<p>主要功能: • 提示注入检测(超过 50 种模式) • 政策引擎(登录/支付阻止) • 设计上故障安全 • 审计日志和请求 ID • Python SDK(同步 + 异步) • RAG 数据清理<p>PyPI: pip install safebrowse<p>我们希望获得来自 AI 基础设施、安全性和代理构建者的反馈。
查看原文
We built SafeBrowse — an open-source prompt-injection firewall for AI systems.<p>Instead of relying on better prompts, SafeBrowse enforces a hard security boundary between untrusted web content and LLMs.<p>It blocks hidden instructions, policy violations, and poisoned data before the AI ever sees it.<p>Features: • Prompt injection detection (50+ patterns) • Policy engine (login&#x2F;payment blocking) • Fail-closed by design • Audit logs &amp; request IDs • Python SDK (sync + async) • RAG sanitization<p>PyPI: pip install safebrowse<p>Looking for feedback from AI infra, security, and agent builders.