Drawbridge

About

Drawbridge protects your personal information before it reaches AI chat sites. Powered by the on-device Rampart model, it detects names, addresses, phone numbers, emails, SSNs, credit cards, and more — then replaces them with stable placeholder tokens like [GIVEN_NAME_1] before your message is sent. The AI site only ever sees the placeholder, while Drawbridge restores the real values in replies right on your device.

How it works:
- As you type, detected PII is highlighted in the composer and a chip shows how many items will be protected.
- When you hit send, the real prompt is swapped for a protected version on the wire.
- When the AI replies, placeholder tokens are restored to real values locally before they reach your eyes.

Key features:
- On-device processing: the 14.7 MB Rampart model runs locally in your browser via WebGPU/WASM. No data sent to any external service for detection.
- Fail-closed by default: if the model can't scan a message in time, sending is blocked rather than letting raw text through.
- Supports claude.ai (first surface; architecture is site-agnostic).

Honest limitations:
- Rampart is ~98.4% recall — a strong first line of defense, not a guarantee.
- Text messages only; attachments, file uploads, and pasted images are not scanned.
- Interception covers fetch only (claude.ai uses fetch today).
- Session tables are in-memory only; after a browser restart, historical placeholders may not restore correctly until a new conversation starts.

Permissions:
- Storage: to save your settings locally.
- Scripting: to inject content scripts on supported sites.
- Offscreen: to host the Rampart model in a dedicated execution context.