Sovereign Shield — LLM PII guard

Redacts Swiss, EU & international identifiers before ChatGPT/Gemini/Claude, then restores them in the reply — 100% local.

Sovereign Shield keeps sensitive data out of web-based AI chat assistants. Before your prompt leaves the browser, it replaces personal identifiers, developer keys, and custom sensitive terms with placeholders — then automatically restores the real values in the AI's reply so nothing sensitive leaves your machine. Everything runs 100% locally: no accounts, no API keys required, no external servers, and no tracking or analytics. How It Works Identifiers are checksum-validated (shape and check digits must match) so ordinary text is untouched and false positives are minimized. Secrets are matched using structured credential patterns. It acts on the message you type — not on files you attach, so redact a document or codebase before uploading it. What It Detects • Personal & Financial Info: Swiss AHV/AVS, global IBANs, credit cards, phone numbers, and email addresses. • National & Health IDs: Official tax, health, and government identity numbers across major regions (including EU, UK, Americas, and Asia). • Developer Secrets: Common API keys, cloud service access tokens, authentication tokens (JWTs), and PEM private keys. • Custom Rules: Your own custom keywords, internal code names, client names, domains, or regular expressions (regex) — plus a built-in library of ready-made rules (US Social Security numbers, UK National Insurance numbers, internal IP addresses, internal hostnames, MAC addresses) you can add with one click, no regex required. See Exactly What You Send A live counter above the chat box shows how many sensitive items will be kept local before you send. Click Inspect to open a side-by-side view of your prompt versus what the AI actually receives, plus a list of every active placeholder and its real value — so you can verify a redaction, remove a false positive, or swap a stand-in on the spot. When you copy the AI's reply, the real values come back too, matching what's on your screen. And if a site changes its internal API so a message would go out uninspected, you're warned instead of left guessing — with a one-click way to report it, so a moved endpoint gets fixed quickly. Optional Smokescreen Mode Send realistic stand-ins (e.g., alice.morgan@example.org instead of [EMAIL_1]) so AI models generate more natural, context-aware responses without seeing your real data. (Applies to emails and custom terms only; IDs, cards, and secrets always use standard placeholders.) Privacy & Control • Complete Granularity: Toggle specific detection categories on or off and define custom privacy rules. • On-Device Activity Log: Monitor redacted items locally (only metadata like type, timestamp, and site are logged — never the raw values). • Zero Storage: The value-to-placeholder map exists solely in your active tab's memory and is erased when closed. Open Source Review the source code on GitHub: https://github.com/acoseac/sovereign-shield