GitHub - privacyshield-ai/privacy-firewall: A local AI-powered DLP solution

4 min read Original article ↗

PrivacyFirewall (Local AI Privacy Shield / Local LLM DLP)

⭐ 100+ stars in 24 hours | 🍴 7 forks | 🔥 Trending

"Solves a major security concern" — Security Professional "Very elegant and impressive solution" — LinkedIn User

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

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

PrivacyFirewall Demo

PrivacyFirewall is a local-first PII and secrets firewall for AI tools like ChatGPT, Claude, and Gemini.It blocks risky paste events, warns as you type, and (optionally) uses a lightweight on-device Transformer model for deeper PII detection.

🔒 **No data ever leaves your machine.**Everything runs locally in your browser or through an optional local API.You can verify this by inspecting the network panel and reading the open-source code.

🚨 Why This Project Exists

Modern AI tools make it extremely easy to leak sensitive information:

  • Emails & phone numbers

  • API keys & credentials

  • Customer or employee data

  • IP & MAC address

  • Internal logs & stack traces

  • Regulated personal information (PII/PHI)

Traditional enterprise DLP tools don’t cover AI chat prompts.

PrivacyFirewall adds a zero-trust privacy shield BEFORE your text ever reaches a third-party AI system.

What PrivacyFirewall gives you:

  • ✋ Human-in-the-loop protection for accidental leaks

  • 🔒 100% local processing (browser + localhost only)

  • ⚡ Practical protection (regex + optional transformer NER)

  • 🧩 Friendly UX (warnings, paste-block modals, override options)

  • 🛠 OSS and auditable (MV3 + FastAPI + Hugging Face stack)

🧠 How It Works

Two Layers of Protection

  1. **Lite Mode (regex-only)**Runs instantly in the extension — no setup needed.

  2. **AI Mode (optional, local LLM)**Uses a local FastAPI agent + transformer model for deeper detection(People, organizations, locations, contextual entities).

High-level architecture

graph TD
    A[User Pastes/Types Text]:::blueNode -->|Intercept| B(Chrome Extension):::blueNode
    B -->|Regex Check| C{Contains Secrets/PII?}
    C -->|Yes & Paste| D[BLOCK & WARN]:::redNode
    C -->|Yes & Typing| E[SHOW WARNING BANNER]:::redNode
    C -->|No| F{Local Engine Online?}
    F -->|No| G[Allow]:::blueNode
    F -->|Yes| H[Python Local Engine]:::blueNode
    H -->|BERT Model| I{AI Detected PII?}
    I -->|Yes & Paste| D
    I -->|Yes & Typing| E
    I -->|No| G

    classDef blueNode fill:#2563eb,stroke:#1e40af,color:#fff
    classDef redNode fill:#dc2626,stroke:#b91c1c,color:#fff
Loading

  • Regex Mode covers secrets quickly

  • AI Mode enhances detection when the local engine is running

  • If the agent goes offline → extension falls back automatically

🚀 Quickstart (Local Development)

Prerequisites

  • Python 3.10+

  • Chrome/Chromium/Edge

  • Git

1) Clone

$ git clone https://github.com/privacyshield-ai/privacy-firewall.git

$ cd privacy-firewall

2) Run the Local Engine (optional for AI Mode)

$ cd src/engine  python -m venv .venv  

$ source .venv/bin/activate       # Windows: .venv\Scripts\activate

$ pip install --upgrade 

$ pip install -r requirements.txt

$ uvicorn main:app --host 127.0.0.1 --port 8765

3) Install the Chrome Extension

  1. Visit: chrome://extensions

  2. Enable Developer mode

  3. Click Load unpacked

  4. Select: src/extension/

You now have Lite Mode running with regex-based detection.

4) Try It Out

Go to:

Paste:

My email is john.doe@example.com   `

→ Paste is intercepted, modal appears.

Paste:

→ Detected as AWS key → blocked.

Enable AI Mode (when popup UI is ready), type:

 Meeting notes from Sarah Thompson at HR...

→ Local transformer flags PERSON → warns you.

🔍 Detection Coverage

Regex Mode (Fast, Offline, Default)

  • Email address

  • Phone number

  • Credit card candidate

  • MAC address

  • IPv4 address

  • AWS access keys

  • JWT tokens

  • Private key blocks

  • Generic API key / hash patterns

  • US SSN (basic pattern)

AI Mode (Local Transformer)

Powered by dslim/bert-base-NER:

  • PERSON

  • ORGANIZATION

  • LOCATION

  • Additional named entities

  • Helpful for ambiguous or context-based leakage

📁 Project Layout

🧪 Development

Run Detection Tests

src/extension/        Chrome MV3 extension (content script, background worker, UI assets)
src/engine/           FastAPI service + transformer model wrapper
src/engine/models/    Model utilities (Hugging Face pipeline)
src/engine/tests/     Basic test harness for detection

Model & Cache Notes

  • HuggingFace models live in ~/.cache/huggingface/

  • Delete this directory to force a fresh download

🛠️ Future Improvements

  • Extension settings UI (enable/disable regex/AI modes)

  • Add per-site allow/deny lists

  • Add secret-type redaction instead of full block

  • Package engine as a binary or desktop app

  • Explore transformer.js for in-browser inference.

  • Automated CI + browser testing

❗ Troubleshooting

“Engine Offline” Banner

  • Ensure the Python engine is running

  • Confirm nothing else uses port 8765

  • Lite mode will still block regex-based secrets

“It didn’t flag a name I typed”

  • Ensure AI Mode is enabled + engine is online

  • NER models are probabilistic; long names work best

  • Confidence threshold is tunable in transformer_detector.py

🤝 Contributing

PRs and issues are welcome!Please include:

  • OS & browser version

  • Reproduction steps

  • Model version (if reporting AI false positives/negatives)

🔐 Security & Privacy Notes

  • No prompts or text ever leave your machine

  • Extension communicates only with:

    • Browser local context

    • Optional localhost API at 127.0.0.1:8765

  • No analytics, telemetry, or external logging

  • Review src/extension/content-script.js and DevTools → Network tabto verify behavior

📄 License

MIT License.See LICENSE for full text.