PrivacyFirewall

👋 If you're trying PrivacyFirewall, please star the repo!

It helps others discover the project and motivates development.
Takes 2 seconds → ⭐ (top right)

Stop AI Data Leaks Before They Happen
100% Local • Zero Server • Full Control

The Problem

Every day, sensitive data gets leaked to AI chatbots:

• 📧 Customer emails pasted into ChatGPT for summarization
• 🔑 API keys accidentally included in code snippets
• 👤 Employee names shared in meeting notes
• 💳 Credit card numbers copied from support tickets
• 🏠 Home addresses in shipping data analysis

Traditional DLP tools don't protect AI chat interfaces. PrivacyFirewall does.

The Solution

PrivacyFirewall intercepts sensitive data before it reaches AI tools — running entirely in your browser with no external servers.

Key Features

How It Works

graph TD
    A[User Pastes/Types Text] -->|Intercept| B(Content Script)
    B -->|Layer 1| C{Regex Scan}
    C -->|Match Found| D{Block or Warn?}
    D -->|Block| E[🛑 Show Modal]
    D -->|Warn| F[⚠️ Show Banner]
    C -->|No Match| G{AI Engine Ready?}
    G -->|Yes| H[ONNX Model in Browser]
    H -->|Entities Found| D
    H -->|Clean| I[✅ Allow]
    G -->|No| I

    style E fill:#dc2626,color:#fff
    style F fill:#f59e0b,color:#fff
    style I fill:#22c55e,color:#fff

Two-Layer Protection

1. Instant Regex Layer — Catches obvious patterns (emails, credit cards, API keys) in milliseconds
2. AI Layer — BERT Named Entity Recognition detects names, organizations, and locations that regex misses

Both layers run 100% locally in your browser. No Python server. No API calls. No cloud.

Quick Start

Option 1: Download & Install (No Code Required)

Perfect for trying it out — takes 30 seconds

1. Download the latest release:

   

2. Unzip the downloaded file

3. Install in Chrome:

   • Go to chrome://extensions
   • Enable Developer mode (toggle in top right)
   • Click Load unpacked
   • Select the unzipped folder

4. Done! Visit ChatGPT and try pasting:

   Contact john.doe@company.com or call 555-123-4567
   

Option 2: Build from Source (For Developers)

Screenshots

Paste Blocked Sensitive data detected with highlighting	Typing Warning Real-time detection as you type
Extension Popup Quick status and settings access	Settings Page Full control over detection rules

Detection Coverage

Regex Detection (Instant)

AI Detection (BERT NER)

Configuration

Settings Page

Access via the extension popup → Open Settings

• PII Rules: Enable/disable detection for each type
• Block vs Warn: Choose blocking modal or warning banner
• Protected Sites: Manage which AI tools are protected
• AI Confidence: Adjust sensitivity threshold (0-100%)
• Behavior: Toggle real-time typing detection

Protected Sites (Default)

• ✅ ChatGPT (chat.openai.com, chatgpt.com)
• ✅ Claude (claude.ai)
• ✅ Gemini (gemini.google.com)
• ✅ Copilot (copilot.microsoft.com)
• ✅ Poe (poe.com)
• ✅ Grok (grok.com)
• ✅ DeepSeek (chat.deepseek.com)

Architecture

src/extension/
├── manifest.json          # Chrome MV3 manifest
├── background.js          # Service worker (message routing)
├── content-script.js      # Page injection (paste/typing interception)
├── offscreen.js           # AI model execution environment
├── offscreen.html         # Offscreen document container
├── lib/
│   └── transformer-detector.js  # BERT NER model wrapper
├── modules/
│   ├── config.js          # Regex patterns & constants
│   ├── scanner.js         # Detection orchestration
│   ├── settings.js        # Chrome storage management
│   ├── event-handlers.js  # Paste & input handlers
│   └── ui/
│       ├── modal.js       # Blocking modal component
│       ├── banner.js      # Warning banner component
│       └── styles.js      # Shadow DOM styles
└── ui/
    ├── popup.html/js/css  # Extension popup
    └── settings.html/js/css # Settings page

Technology Stack

• Extension: Chrome Manifest V3, ES Modules
• AI Runtime: ONNX Runtime Web (WASM)
• Model: Xenova/bert-base-NER-uncased via Hugging Face Transformers.js
• UI Isolation: Shadow DOM (no CSS conflicts with host pages)

Privacy & Security

What We DON'T Do

• ❌ Send data to external servers
• ❌ Log or store your text
• ❌ Use analytics or telemetry
• ❌ Make any network requests (except model download on first run)

What We DO

• ✅ Process everything locally in your browser
• ✅ Cache the AI model locally after first download
• ✅ Store settings in Chrome's encrypted sync storage
• ✅ Provide fully auditable open-source code

Verify yourself: Open DevTools → Network tab. You'll see zero outbound requests during detection.

Development

Build

cd src/extension
npm install
node build.js

Run Tests

node tests/content-script.test.js

Project Requirements

• Chrome 120+ (for Offscreen Documents API)
• Node.js 18+ (for building)

Troubleshooting

"AI Model Loading..." stays forever

• Check DevTools console for errors
• The model (~50MB) downloads on first run — may take a minute on slow connections
• Try clearing extension data and reloading

Extension doesn't detect on some sites

• Check if the site is in your protected sites list
• Some sites use iframes — detection may be limited
• Open an issue with the site URL

False positives/negatives

• Adjust AI confidence threshold in settings
• Some patterns (like short names) may not be detected
• Report edge cases as issues

Roadmap

• Firefox/Safari support
• Custom regex patterns via settings
• Redaction mode (replace vs block)
• Export/import settings
• Keyboard shortcuts
• Enterprise policy support

Contributing

PRs welcome! Please include:

• Browser version
• Steps to reproduce
• Expected vs actual behavior

License

MIT License — see LICENSE

Built for privacy. Runs locally. Open source.
Report Bug • Request Feature