BiVelio Shield — private prompts for ChatGPT & Claude

Keep your private data out of the AI cloud — without changing how you work. BiVelio Shield sits between you and ChatGPT or Claude and removes sensitive information from the text you type on your own machine, before it leaves your browser (on a paid plan with a names gateway, each message also goes to that gateway first, with every value the extension detects already masked — see below). The AI provider receives placeholder tokens in place of those values — never the real ones — and the extension puts the real values back into the reply on screen, so your conversation reads normally. Verified end to end on both sites. On ChatGPT: a prompt carrying an email reaches the provider as a [BV:…] token, not the raw value — reload the thread and only the token was ever stored, and the reply shows your real values restored on screen. On Claude: the same round-trip was run on 2026-08-14 with a control in the same thread. IBANs, cards, national IDs and API keys run through the same engine and are covered by its test suite. What it protects (free, on-device, no signup): • Credit-card numbers (checksum-validated) • IBANs and bank details • National IDs — Spanish DNI/NIE/CIF, US SSN, and more • Emails and phone numbers — French numbers in their national format (06 12 34 56 78, also with dots or dashes, or run together after a label like Tél.), and Andorran numbers, too • API keys & secrets — OpenAI, AWS, GitHub, Stripe, Slack, and others • IP addresses • Postal addresses that name a street type (calle, carrer, rue, rua…) in Spanish, Catalan, French or Portuguese, and Spanish cadastral references • Passwords written in Spanish or Catalan, passport numbers, vehicle number plates and dates of birth, when a label names them — on the same line, on the line above or in a table's header row • Health, religious, biometric and genetic terms — a curated multilingual lexicon (GDPR Art. 9 special categories) • Your own confidential terms — client, matter or project names you add in the extension's options page. The list stays on your device, stored as you typed it (not encrypted), and is never sent Optional — "Redact on shape": turn it on to also hide mistyped IDs that fail their checksum. Off by default, and we tell you plainly it may over-redact things like order numbers or tracking codes — your choice. And even with it off: a checksum-valid identifier can occasionally swallow a word or two next to it, so the provider sees less of your sentence than you typed — never more. Honest about what it does NOT do — no dark patterns: • It does not remove people's names or general free-form text on its own, on any tier (no fixed shape to match), and an Art. 9 phrase written in wording its curated lexicon does not list will also pass. Names are looked for on one route only, and only for an organisation on a paid plan. With the Team plan that route comes with the plan: the address of the names service BiVelio operates arrives signed inside your organisation's own licence, so there is nothing to install and no URL to type, BiVelio acts as your organisation's processor under a data-processing agreement for it, and on Team an owner or an administrator can switch the layer off for everyone from the dashboard — while on Pro, and on any install whose gateway comes from the central managed policy, that policy takes precedence and the way off is to remove the address from it. On any paid plan your administrator can point the extension at a gateway of their own from that central managed policy. Either way the text of each message (up to 20,000 characters on Pro and 50,000 on Team, sent in pieces of up to 4,000 — and on Team also the text of an attachment, sent after the local redaction has replaced its identifiers) goes to that gateway before the local redaction runs — with the terms on your own or your organisation's confidential-terms list, and every value the extension detects on its own (identifiers, account and card numbers, contact details, secrets), first replaced by asterisks — and comes back as the positions of the personal data the gateway detects — names and places, and also contact details, identifiers, financial data, network addresses, secrets and GDPR Art. 9 special-category terms. In ChatGPT it also fires when you paste 1,000 characters or more into the chat box, at the moment of the paste, so a message you then decide not to send can still reach the gateway. It is a layer that reduces the leak of names and, measured, does not eliminate it (4.875%). A long message also meets a time limit: at the names service's measured speed, only about 45–75% of a 50,000-character message is checked for names before it runs out; the rest is sent without that check — still redacted locally — and a card says the message was not fully checked. With neither of those two the extension looks for names on no tier. • Voice / realtime mode is not protected. It redacts the text you type; a spoken turn in ChatGPT Advanced Voice or Claude voice travels over a separate live-audio connection it never sees — and a text tool cannot redact audio. Type, don't speak, when a turn carries sensitive data. • Claude's "Cowork" composer is not covered. The extension redacts Claude's normal chat send; Cowork posts to a different endpoint that it deliberately does not touch. Treat anything typed into Cowork as unprotected. • Files and attachments are inspected only on the Team tier. On the Free and Pro tiers the engine is text-only: an uploaded document, PDF, or a photographed passport is sent to the provider as-is, and a notice tells you when that happens. Team redacts them in your browser before they upload: text files, PDFs that have a text layer, and modern Word, Excel, PowerPoint and OpenDocument files. That includes the values typed into a PDF form's fields, OpenDocument files saved by LibreOffice, and an identifier split across several text cells of a table. Older pre-2007 binary Office documents cannot be read by any browser extension and are NOT…