-
-
Live on ChatGPT: Veil caught 6 private values before ChatGPT saw the paste
-
Install once, always on — Veil checks every paste into ChatGPT, Gemini and Claude
-
A Chrome extension that lives inside ChatGPT, Gemini and Claude
-
Before vs after: what ChatGPT receives without Veil (raw PAN, Aadhaar) and with Veil (tags only)
-
The answer comes back and Veil restores the real values — only on your screen
-
Live on Claude: Veil caught 6 values before Claude saw them
-
Claude's own reply: "the client's details are masked"
-
Under the hood: rule packs, Gemini Nano on-device, Guess Who attacker, cloud Gemini on tags only
-
Guess Who: Veil attacks itself on-device — could a stranger work out who this is?
-
Benchmark: 110 of 113 adversarial documents with zero leaks (98.5% vs 69.7% rules-only)
-
Veil — the cloud gets the thinking. Never your secrets.
Inspiration
Accountants, lawyers, HR teams and doctors hold other people's secrets — and paste them into ChatGPT, Gemini and Claude every day. In one 2026 survey, 79% of workers admitted pasting sensitive data into AI tools. In India, the DPDP Rules start to bite on 13 May 2027, with penalties of up to ₹250 crore. Cloud models reason better; on-device models are private. Hybrid inference finally lets one app use each for what it's good at.
What it does
Veil is a Chrome extension that lives inside ChatGPT, Gemini and Claude. Install it once and it's always on: every time you paste into one of those sites, Veil checks it first — on your device — before the website sees a single character.
- Rule packs (India, EU, UK, US) catch IDs instantly — PAN, Aadhaar, GSTIN, IFSC, UPI, IBAN, cards, phone numbers — with checksums and unicode normalisation.
- Gemini Nano, on your device, reads the text like a person would: names, places, and quiet giveaways like "the only female cardiologist in Ratnagiri".
- Guess Who — Veil attacks itself: a second on-device pass asks "could a stranger work out who this is?" and whatever evidence it cites gets masked too.
- Values become smart tags —
[PAN_1],[AMOUNT_1 · ₹15–20L]— so the AI can still reason about tax slabs and deadlines. A fail-closed leak gate re-checks the exact payload. - The AI answers with tags; Veil restores the real values on your screen only. ChatGPT, Gemini and Claude never have them. In our live demo, Claude itself replied: "the client's details are masked."
Veil also ships as Veil Studio (web app with a privacy receipt — a SHA-256 of exactly what was sent), a playable Beat Veil game and a Privacy Lens on the landing page, and an on-device benchmark page.
How we built it — Firebase AI Logic hybrid inference
getGenerativeModel(ai, { mode: InferenceMode.ONLY_ON_DEVICE })with JSON-schemaresponseConstraintfor three on-device tasks (names/places, indirect clues, attacker). Sensitive work is deliberately never allowed to fall back to the cloud.getGenerativeModel(ai, { mode: InferenceMode.ONLY_IN_CLOUD })for reasoning over masked text only;response.inferenceSourceis surfaced per stage.- In the extension, Gemini Nano runs in an offscreen document through the same Firebase AI Logic code.
- App Check (reCAPTCHA Enterprise), Authentication (Google), Cloud Firestore (receipt history — security rules accept only counts and a 64-hex hash), Remote Config (cloud model, timeouts, never-mask terms), Hosting.
- React 19 + Vite + TypeScript + Tailwind 4 + Motion; 82 Vitest (unit, property-based, security) + 21 Playwright tests, including tests that load the real extension.
Challenges we ran into
- Small on-device models over-mask and paraphrase. We added a merge layer (rules win, protected terms like
Form 26AS/AY 2026-27are never masked, fuzzy re-location of paraphrased spans) and measured utility, not just recall. - The attacker's self-reported scores were meaningless, so risk is computed from verified evidence instead.
- Streaming chat UIs split
[PERSON_1]across elements; the extension restores values per block without breaking the page's DOM. - Intercepting a paste before the site's own handlers see it — across three very different editors (textarea, contenteditable, ProseMirror).
Accomplishments that we're proud of
- 98.5% of private items never reached the cloud across a 113-case adversarial benchmark (typos, Hinglish, hidden IDs, CSV/JSON, quasi-identifiers) — vs 69.7% for rules alone. 110 / 113 documents with zero leaks; all 3 leaks happened during an on-device timeout, which Veil flags for review.
- Works live on ChatGPT, Gemini and Claude; detection keeps working offline.
- Verified in DevTools that the cloud payload contains zero personal values.
What we learned
Hybrid isn't "on-device as a fallback" — it's a division of labour. Keep anything that sees raw secrets strictly on-device, send the cloud only what it needs to think, and prove it with a receipt.
What's next
Android with Gemini Nano via ML Kit + Firebase AI Logic, more region packs via Remote Config, team policies, and publishing to the Chrome Web Store.
Built on research
PAPILLON (arXiv:2410.17127) · Hide and Seek (arXiv:2309.03057) · Language Models are Advanced Anonymizers (ICLR 2025, arXiv:2402.13846) · Beyond Memorization (ICLR 2024, arXiv:2310.07298)
Built With
- chrome
- cloud-firestore
- fast-check
- firebase-ai-logic
- firebase-app-check
- firebase-authentication
- firebase-hosting
- firebase-remote-config
- gemini
- gemini-nano
- javascript
- motion
- playwright
- react
- recaptcha-enterprise
- tailwindcss
- typescript
- vite
- vitest