锌的故事
As autonomous agents execute tools, modify codebases, and interact with the open internet, zn acts as the sacrificial boundary between untrusted inputs and agent runtime.
我们为什么构建 zn
AI Agent 正在成为人与系统之间的新界面。它们代表我们阅读邮件、管理日程、执行代码并调用 API。但能力越大，风险也越大。
提示词注入攻击利用了 LLM 的基础架构。攻击者可以在看似无害的数据（日历邀请、邮件、文档）中嵌入恶意指令，劫持你的 Agent。
我们构建 zn，是因为 Agent 值得更好的防护。正如钢铁上的镀锌层，zn 提供牺牲式防护——我们替你承受冲击，让你的系统保持安全。
Our Engineering Principles
Zero External Dependencies & Local-First
Security tooling should never expand your attack surface. zn-gate runs 100% offline with zero external npm dependencies, weighing under 15 kB and deciding in sub-millisecond speeds (< 0.1 ms) directly on your local CPU.
Full-Lifecycle Bidirectional Coverage
Protection must extend beyond the user prompt. zn inspects incoming user inputs (analyze_prompt), pre-flight tool call arguments (check_tool_call), and post-execution third-party outputs (check_tool_result) against Indirect Prompt Injection.
Deterministic Speed + Deep Neural Camera
Immediate regex pre-normalization (homoglyphs, zero-width chars, inline comments, Base64 decoding) catches active evasions instantaneously, fused with an optional cloud ONNX INT8 neural gate (99.4% precision) for complex zero-day vectors.
Open Science & Reproducibility
No security through obscurity. We openly publish our 23.7k benchmark dataset on Hugging Face, our gate package on npm, and our code on GitHub. Every performance claim is independently verifiable.
Supported by Global Cloud Programs
AWS for Startups (Activate)
zn 是 AWS for Startups 计划的一部分，为我们提供 Activate 积分、架构指导与初创优惠——从边缘计算带来的亚毫秒延迟，到企业级安全与合规。
Universal MCP Architecture
Designed from day one for the Model Context Protocol (MCP). Out-of-the-box native compatibility with Cursor, Claude Code, Antigravity, OpenCode, Codex, Hermes Agent, and custom LLM workflows.
我们的防护对象
设想一个运维 Agent 正在读取一张包含运行指令的支持工单 rm -rf /，zn 为 Agent 提供牺牲式边界与取证轨迹，而不是把每段字符串都当作指令。
| 方案 | 结果 |
|---|---|
| zn security gate | 拦截被注入的工具调用并保留取证元数据 |
| 传统 WAF | 可能看到传输特征，但无法理解 Agent 的语义意图 |
| 无防护 | Agent 在没有安全层的情况下直接决定并执行 |
Security & Responsible Disclosure (RFC 9116)
We practice coordinated vulnerability disclosure. If you discover a potential security flaw, guardrail bypass, or evasion vulnerability in zn-gate or our cloud infrastructure, please report it directly to our security team. We provide safe harbor for researchers acting in good faith.
保持联系
Platform integrations, partnership inquiries, and product questions.
Vulnerability reports, penetration testing findings, and security audits.
Direct contact with lead maintainer tljohnsilver for architectural discussions.
准备好保护你的 Agent 了吗？
免费开始 48 小时 Hobby 试用。无需信用卡。