kocaemre/zeguard
GitHub: kocaemre/zeguard
ZeGuard 是一款运行在浏览器端的安全工具，用于检测 PDF、DOCX 和 PPTX 文档中隐藏的 Prompt Injection 攻击技术。
Stars: 0 | Forks: 0
# user wants a translation of a Markdown document from Turkish to Simplified Chinese.
**Constraints Checklist & Confidence Score:**
1. Keep code placeholders (`
```
git clone https://github.com/kocaemre/zeguard.git
cd zeguard
npm install
npm run dev
```
`)? Yes.
2. Keep HTML image placeholders (`
`)? Yes.
3. Keep link URL placeholders (`https://docguard.zepatrol.xyz`)? Yes.
4. Translate link text? Yes.
5. Preserve HTML blocks? Yes.
6. Keep Markdown structure? Yes.
7. Translate headings? Yes.
8. Preserve technical terms (English)? Yes.
Confidence Score: 5/5
**Mental Sandbox Simulation:**
- Title: `# 🛡️ ZeGuard — Gizli Prompt Injection Dedektörü` -> `# 🛡️ ZeGuard — 隐蔽 Prompt Injection 检测器` (ZeGuard is a name, Prompt Injection is technical).
- Demo: `[docguard.zepatrol.xyz](https://docguard.zepatrol.xyz)` -> `[docguard.zepatrol.xyz](https://docguard.zepatrol.xyz)` (Link text is a domain, usually kept as is or translated if it's a description. Here it looks like the URL itself as text. I will keep it as is to be safe, or translate if it's clearly a label. The rule says "translate the link text". "docguard.zepatrol.xyz" is a domain. I will keep the domain as is because it's an identifier, but if it were "Demo Site", I'd translate. Wait, the rule says "Example: [Overview]... -> [概述]". If the text is a URL, it's safer to keep it, but strictly following "translate link text" might imply making it Chinese. However, usually, domains are not translated. I will keep the domain text as is because it serves as the address.)
- Heading: `## 🎯 Ne Yapar?` -> `## 🎯 功能` (What does it do? -> Functionality).
- Table:
- `Teknik` -> `技术`
- `Açıklama` -> `描述`
- `Invisible Font` -> `Invisible Font` (Technical term/Proper noun concept).
- `Font rengi ≈ arka plan rengi (Delta-E ≤ 10) kontrolü` -> `字体颜色 ≈ 背景颜色 (Delta-E ≤ 10) 检查`
- `Tiny Font` -> `Tiny Font`
- `4pt altı metin — gözle görünmez, LLM okur` -> `小于 4pt 的文本 — 肉眼不可见，但 LLM 可读`
- `Hidden Text` -> `Hidden Text`
- `DOCX w:vanish ve w:webHidden XML tag'leri` -> `DOCX w:vanish 和 w:webHidden XML 标签`
- `Unicode Invisible` -> `Unicode Invisible`
- `8 farklı invisible Unicode karakteri (özellikle 3+ ardışık)` -> `8 种不同的不可见 Unicode 字符（特别是 3 个以上连续字符）`
- `Metadata Injection` -> `Metadata Injection`
- `docProps/core.xml ve custom.xml içinde injection pattern'ları` -> `docProps/core.xml 和 custom.xml 内部的注入模式`
- `Pattern Matching` -> `Pattern Matching`
- `50+ Türkçe ve İngilizce prompt injection pattern veritabanı` -> `50+ 土耳其语和英语 prompt injection 模式数据库`
- Privacy: `Tüm analiz **tarayıcıda** gerçekleşir.` -> `所有分析均在**浏览器**中完成。`
- Tech Stack:
- `React 18` -> `React 18`
- `Vite — build tool` -> `Vite — 构建工具`
- `Tailwind CSS — stil` -> `Tailwind CSS — 样式`
- `pdf.js — PDF parsing` -> `pdf.js — PDF 解析`
- `mammoth.js — DOCX parsing` -> `mammoth.js — DOCX 解析`
- `PptxGenJS — PPTX parsing` -> `PptxGenJS — PPTX 解析`
- Installation: `## 🚀 Kurulum` -> `## 🚀 安装`
- Build: `### Build` -> `### 构建`
- Project Structure: `## 📁 Proje Yapısı` -> `## 📁 项目结构`
- Developer: `## 👨💻 Geliştirici` -> `## 👨💻 开发者`
- License: `## 📄 Lisans` -> `## 📄 许可证`
**Refinement on Technical Terms:**
- "Prompt Injection" -> Keep as "Prompt Injection" (Rule 9).
- "Invisible Font" -> Keep as "Invisible Font" or "隐形字体"? Rule 9 says "Domain-specific terminology widely known in English... e.g. token, prompt...". "Invisible Font" is a specific attack technique name. I will keep it in English to be safe and professional, or translate the description. The table has "Teknik" (Technique) as the header. I will keep the technique names in English as they are likely standard terms in this security domain.
- "Metadata Injection" -> Keep as "Metadata Injection".
- "Pattern Matching" -> Keep as "Pattern Matching".
**Refinement on Link Text:**
- `[docguard.zepatrol.xyz](https://docguard.zepatrol.xyz)`: The text is the URL itself. I will leave it as is.
- `[0xemrek](https://github.com/kocaemre)`: This is a username/handle. I will leave it as is.
**Final Polish:**
Ensure the tone is professional and the Chinese is natural.
Let's generate the output.
# 🛡️ ZeGuard — 隐蔽 Prompt Injection 检测器
**演示：** [docguard.zepatrol.xyz](https://docguard.zepatrol.xyz)
## 🎯 功能
当您向 AI 系统上传文件时，文档中隐藏的指令可能会操纵 LLM。ZeGuard 会自动检测这些技术。
### 检测到的攻击技术
| 技术 | 描述 |
|--------|----------|
| **Invisible Font** | 字体颜色 ≈ 背景颜色 (Delta-E ≤ 10) 检查 |
| **Tiny Font** | 小于 4pt 的文本 — 肉眼不可见，但 LLM 可读 |
| **Hidden Text** | DOCX `w:vanish` 和 `w:webHidden` XML 标签 |
| **Unicode Invisible** | 8 种不同的不可见 Unicode 字符（特别是 3 个以上连续字符） |
| **Metadata Injection** | `docProps/core.xml` 和 `custom.xml` 内部的注入模式 |
| **Pattern Matching** | 50+ 土耳其语和英语 prompt injection 模式数据库 |
## 🔒 隐私
所有分析均在**浏览器**中完成。您的文件不会发送到任何服务器。
## 🛠️ 技术
- **React 18** + TypeScript
- **Vite** — 构建工具
- **Tailwind CSS** — 样式
- **pdf.js** — PDF 解析
- **mammoth.js** — DOCX 解析
- **PptxGenJS** — PPTX 解析
## 🚀 安装
```
git clone https://github.com/kocaemre/zeguard.git
cd zeguard
npm install
npm run dev
```
### 构建
```
npm run build
# dist/ klasörü oluşur, herhangi bir static hosting'e deploy edilebilir
```
## 📁 项目结构
```
zeguard/
├── src/
│ ├── analyzers/ # Analiz modülleri
│ │ ├── color-analyzer.ts # Invisible font tespiti
│ │ ├── font-analyzer.ts # Tiny font tespiti
│ │ ├── metadata-analyzer.ts # Metadata injection
│ │ ├── pattern-analyzer.ts # Pattern matching
│ │ └── unicode-analyzer.ts # Invisible unicode
│ ├── parsers/ # Dosya ayrıştırıcıları
│ │ ├── pdf-parser.ts
│ │ ├── docx-parser.ts
│ │ └── pptx-parser.ts
│ ├── ui/ # React bileşenleri
│ │ ├── DropZone.tsx
│ │ ├── ResultsPanel.tsx
│ │ ├── FindingCard.tsx
│ │ ├── ScanProgress.tsx
│ │ └── AttacksPage.tsx
│ ├── patterns/
│ │ ├── tr.json # Türkçe injection pattern'ları
│ │ └── en.json # İngilizce injection pattern'ları
│ ├── reporter/
│ │ └── report-builder.ts
│ └── App.tsx
├── public/
├── index.html
└── vite.config.ts
```
## 👨💻 开发者
**Emre Koca** — [0xemrek](https://github.com/kocaemre)
## 📄 许可证
MIT
`)? Yes.
3. Keep link URL placeholders (`https://docguard.zepatrol.xyz`)? Yes.
4. Translate link text? Yes.
5. Preserve HTML blocks? Yes.
6. Keep Markdown structure? Yes.
7. Translate headings? Yes.
8. Preserve technical terms (English)? Yes.
Confidence Score: 5/5
**Mental Sandbox Simulation:**
- Title: `# 🛡️ ZeGuard — Gizli Prompt Injection Dedektörü` -> `# 🛡️ ZeGuard — 隐蔽 Prompt Injection 检测器` (ZeGuard is a name, Prompt Injection is technical).
- Demo: `[docguard.zepatrol.xyz](https://docguard.zepatrol.xyz)` -> `[docguard.zepatrol.xyz](https://docguard.zepatrol.xyz)` (Link text is a domain, usually kept as is or translated if it's a description. Here it looks like the URL itself as text. I will keep it as is to be safe, or translate if it's clearly a label. The rule says "translate the link text". "docguard.zepatrol.xyz" is a domain. I will keep the domain as is because it's an identifier, but if it were "Demo Site", I'd translate. Wait, the rule says "Example: [Overview]... -> [概述]". If the text is a URL, it's safer to keep it, but strictly following "translate link text" might imply making it Chinese. However, usually, domains are not translated. I will keep the domain text as is because it serves as the address.)
- Heading: `## 🎯 Ne Yapar?` -> `## 🎯 功能` (What does it do? -> Functionality).
- Table:
- `Teknik` -> `技术`
- `Açıklama` -> `描述`
- `Invisible Font` -> `Invisible Font` (Technical term/Proper noun concept).
- `Font rengi ≈ arka plan rengi (Delta-E ≤ 10) kontrolü` -> `字体颜色 ≈ 背景颜色 (Delta-E ≤ 10) 检查`
- `Tiny Font` -> `Tiny Font`
- `4pt altı metin — gözle görünmez, LLM okur` -> `小于 4pt 的文本 — 肉眼不可见，但 LLM 可读`
- `Hidden Text` -> `Hidden Text`
- `DOCX w:vanish ve w:webHidden XML tag'leri` -> `DOCX w:vanish 和 w:webHidden XML 标签`
- `Unicode Invisible` -> `Unicode Invisible`
- `8 farklı invisible Unicode karakteri (özellikle 3+ ardışık)` -> `8 种不同的不可见 Unicode 字符（特别是 3 个以上连续字符）`
- `Metadata Injection` -> `Metadata Injection`
- `docProps/core.xml ve custom.xml içinde injection pattern'ları` -> `docProps/core.xml 和 custom.xml 内部的注入模式`
- `Pattern Matching` -> `Pattern Matching`
- `50+ Türkçe ve İngilizce prompt injection pattern veritabanı` -> `50+ 土耳其语和英语 prompt injection 模式数据库`
- Privacy: `Tüm analiz **tarayıcıda** gerçekleşir.` -> `所有分析均在**浏览器**中完成。`
- Tech Stack:
- `React 18` -> `React 18`
- `Vite — build tool` -> `Vite — 构建工具`
- `Tailwind CSS — stil` -> `Tailwind CSS — 样式`
- `pdf.js — PDF parsing` -> `pdf.js — PDF 解析`
- `mammoth.js — DOCX parsing` -> `mammoth.js — DOCX 解析`
- `PptxGenJS — PPTX parsing` -> `PptxGenJS — PPTX 解析`
- Installation: `## 🚀 Kurulum` -> `## 🚀 安装`
- Build: `### Build` -> `### 构建`
- Project Structure: `## 📁 Proje Yapısı` -> `## 📁 项目结构`
- Developer: `## 👨💻 Geliştirici` -> `## 👨💻 开发者`
- License: `## 📄 Lisans` -> `## 📄 许可证`
**Refinement on Technical Terms:**
- "Prompt Injection" -> Keep as "Prompt Injection" (Rule 9).
- "Invisible Font" -> Keep as "Invisible Font" or "隐形字体"? Rule 9 says "Domain-specific terminology widely known in English... e.g. token, prompt...". "Invisible Font" is a specific attack technique name. I will keep it in English to be safe and professional, or translate the description. The table has "Teknik" (Technique) as the header. I will keep the technique names in English as they are likely standard terms in this security domain.
- "Metadata Injection" -> Keep as "Metadata Injection".
- "Pattern Matching" -> Keep as "Pattern Matching".
**Refinement on Link Text:**
- `[docguard.zepatrol.xyz](https://docguard.zepatrol.xyz)`: The text is the URL itself. I will leave it as is.
- `[0xemrek](https://github.com/kocaemre)`: This is a username/handle. I will leave it as is.
**Final Polish:**
Ensure the tone is professional and the Chinese is natural.
Let's generate the output.
# 🛡️ ZeGuard — 隐蔽 Prompt Injection 检测器
**演示：** [docguard.zepatrol.xyz](https://docguard.zepatrol.xyz)
## 🎯 功能
当您向 AI 系统上传文件时，文档中隐藏的指令可能会操纵 LLM。ZeGuard 会自动检测这些技术。
### 检测到的攻击技术
| 技术 | 描述 |
|--------|----------|
| **Invisible Font** | 字体颜色 ≈ 背景颜色 (Delta-E ≤ 10) 检查 |
| **Tiny Font** | 小于 4pt 的文本 — 肉眼不可见，但 LLM 可读 |
| **Hidden Text** | DOCX `w:vanish` 和 `w:webHidden` XML 标签 |
| **Unicode Invisible** | 8 种不同的不可见 Unicode 字符（特别是 3 个以上连续字符） |
| **Metadata Injection** | `docProps/core.xml` 和 `custom.xml` 内部的注入模式 |
| **Pattern Matching** | 50+ 土耳其语和英语 prompt injection 模式数据库 |
## 🔒 隐私
所有分析均在**浏览器**中完成。您的文件不会发送到任何服务器。
## 🛠️ 技术
- **React 18** + TypeScript
- **Vite** — 构建工具
- **Tailwind CSS** — 样式
- **pdf.js** — PDF 解析
- **mammoth.js** — DOCX 解析
- **PptxGenJS** — PPTX 解析
## 🚀 安装
```
git clone https://github.com/kocaemre/zeguard.git
cd zeguard
npm install
npm run dev
```
### 构建
```
npm run build
# dist/ klasörü oluşur, herhangi bir static hosting'e deploy edilebilir
```
## 📁 项目结构
```
zeguard/
├── src/
│ ├── analyzers/ # Analiz modülleri
│ │ ├── color-analyzer.ts # Invisible font tespiti
│ │ ├── font-analyzer.ts # Tiny font tespiti
│ │ ├── metadata-analyzer.ts # Metadata injection
│ │ ├── pattern-analyzer.ts # Pattern matching
│ │ └── unicode-analyzer.ts # Invisible unicode
│ ├── parsers/ # Dosya ayrıştırıcıları
│ │ ├── pdf-parser.ts
│ │ ├── docx-parser.ts
│ │ └── pptx-parser.ts
│ ├── ui/ # React bileşenleri
│ │ ├── DropZone.tsx
│ │ ├── ResultsPanel.tsx
│ │ ├── FindingCard.tsx
│ │ ├── ScanProgress.tsx
│ │ └── AttacksPage.tsx
│ ├── patterns/
│ │ ├── tr.json # Türkçe injection pattern'ları
│ │ └── en.json # İngilizce injection pattern'ları
│ ├── reporter/
│ │ └── report-builder.ts
│ └── App.tsx
├── public/
├── index.html
└── vite.config.ts
```
## 👨💻 开发者
**Emre Koca** — [0xemrek](https://github.com/kocaemre)
## 📄 许可证
MIT标签：AI安全, Chat Copilot, DOCX安全, mammoth.js, pdf.js, PDF安全, PPTX安全, React 18, Tailwind CSS, TLS, TypeScript, Unicode隐藏字符, Vite, ZeGuard, 元数据注入, 前端安全, 安全插件, 客户端安全工具, 提示词注入检测, 文件上传安全, 本地文件分析, 浏览器端解析, 网络安全, 自定义脚本, 防御工具, 隐形字体检测, 隐私保护, 默认DNS解析器