Velum Lite — Redact & Mask PII for ChatGPT, Claude 作成者: Velum Privacy

Velum Lite masks personal and sensitive data before you paste it into an AI
assistant, then restores the originals in the reply. Everything runs on your
machine. The extension makes no internet requests and bundles its detection
model locally, so it works fully offline.

How it works
- Mask: paste text into the popup and it masks on the spot. A black bar sweeps
across each identifier, leaves a token behind, and the result lands on your
clipboard. Paste that into the chatbot.
- Unmask: paste the reply back, press Unmask, and your originals are restored
from an in-memory mapping that never leaves the browser.
- Session mapping: see exactly what was taken out and what replaced it. Each
original stays under a bar until you lift it.
- Files: drop a document onto the popup and get a masked Markdown file back.
Common document, text and spreadsheet formats are supported, and optional
offline OCR handles scanned PDFs.
- Optional in-page mode: turn on In-page masking to get a Mask button in the
composer on claude.ai, chatgpt.com and gemini.google.com. Off by default.
Enabling it asks for permission for those three sites only.

What it detects

Contact and financial identifiers run always, in every language and country.
That covers email addresses and phone numbers, bank account and payment card
numbers, passport numbers, and the network and device identifiers that show up
in logs and config files.

Credentials are matched by around fifty built-in patterns spanning cloud
platforms, code hosting, package registries, payment processors, messaging
services and AI providers, together with signed web tokens, private keys,
database connection strings and other high-entropy secrets.

National identifiers are matched by country packs with checksum validation, so
a number that fails its own check digit is left alone instead of masked. Packs
ship today for Spain, the United States, the United Kingdom, France, Germany,
Türkiye and Italy, and each one also carries the address and phone recognisers
for that country. The active set is detected from your browser language and can
be changed in Settings.

People, organizations and places are found by a local multilingual model rather
than a pattern, so unusual names are still caught. They become tokens by
default, or format-valid surrogates if you opt in.

Settings
- Countries and regions: pick which national-identifier packs run, or leave it
on auto.
- Detection sensitivity: how sure the model has to be before it masks a name,
organization or place. Lower catches more and over-masks more.
- Realistic surrogates instead of tokens: replaces detected values with
format-valid fakes rather than opaque markers. Off by default, because a
surrogate that reads as real is easy to forget to unmask.
- OCR scanned PDFs: runs offline OCR on image-only documents. Off by default;
it is slow and adds nothing unless you work with scans.
- Theme: follows your browser, or pin light or dark.
- Use Desktop engine: routes masking to the Velum Desktop app on the same
machine over the loopback interface. Off by default, and only loopback
addresses are accepted.

Privacy
- 100% on device. No servers, no analytics, no accounts.
- The mapping that reverses masking lives in session memory and clears when the
browser closes.
- No host permissions are requested unless you turn on in-page masking.

Free and paid
- Masking and unmasking text in the popup is free, with no trial period and no
expiry. Full detection is included: national identifiers, credentials, and
names, organizations and places through the local model.
- A licence unlocks masking documents and the in-page Mask button. Licences are
sold at velumprivacy.com and the same key also activates Velum Desktop.
- Unmasking is never restricted, licence or not. Anything you have masked, you
can always restore.

Worth knowing
- Detection is strong but not perfect. Read the masked text before you send it.
- Always unmask the reply before forwarding it, above all with realistic
surrogates turned on.