What happens to your data
Trust is the whole point of a tool that hides personal data. So here, without spin, is what happens locally, what leaves your device and what we don’t know.
In short
- Clean before AI runs entirely in your browser — text and files are not sent anywhere, not even to us.
- The AI (ChatGPT, Claude …) only gets the cleaned text with labels, and only if you copy or send it yourself.
- The key that links labels to real data stays with you — in the extension only in the browser’s session memory.
- No account, no cookies, no analytics or tracking.
Where your data goes
Client John Smith, SSN 123-45-6789, IBAN GB82 WEST 1234 5698 7654 32
Client [PERSON_1], SSN [SSN_1], IBAN [IBAN_1]
The label key stays with you onlyUnderstands the meaning, never sees the data. Answers with labels:
Dear [PERSON_1], the refund to [IBAN_1] is confirmed.
Dear John Smith, the refund to GB82 WEST 1234 5698 7654 32 is confirmed.
1. What is processed locally
- Detection of personal data: rules with checksums (national IDs, tax numbers, IBAN …), name dictionaries and grammatical cases — all in JavaScript in your browser.
- A local AI model (our own, ~53 MB, 17 languages) finds names, addresses and companies the rules don’t know. In the extension it is bundled and works offline; in the web tool it downloads once from hypervision.si when you switch it on (~67 MB including the runtime) and is stored in your browser. Only the model is downloaded — never your text.
- Reading PDF, Word and OCR of scans, and redacting PDFs — locally.
- Putting real data back into the AI’s answer — locally, with the key that stays with you.
2. What leaves your device
- Only the cleaned text you copy into an AI yourself or send with a button (for short texts a link to the chatbot opens with the cleaned text).
- In the extension: when you paste into a supported chatbot, the extension intercepts the paste before the page gets it and inserts the cleaned version. Only what you send goes to the AI’s servers.
- Nothing else: the tool does not call any Hypervision server.
3. What is stored
- Web tool: nothing on a server. Known names (e.g. colleagues, clients) are stored in your browser only if you choose so. You can download the key as a .json file.
- Extension: the key and recent answers live in chrome.storage.session — in memory, deleted when the browser closes or with a button. Settings and the known-names list are in chrome.storage.local on your computer; an organisation can distribute the list through a Chrome policy.
- In the extension real data is shown only in the side panel — it is never written into the chatbot page.
4. Extension permissions and why
storageSaving settings locally and the key in session memory.
sidePanelThe side panel where you see real data and the restored answer — separate from the AI page.
offscreenAn invisible extension document that runs the local AI model (WebAssembly). It has no access to web pages or the network.
Access to 11 AI sitesOnly ChatGPT, Claude, Gemini, Microsoft Copilot, Perplexity, DeepSeek, Mistral Vibe, Grok, Poe, Qwen, HuggingChat: to intercept a paste before the page receives it. The extension does not run on other sites.
No networkThe extension has no permission to make web requests and sends nothing — not even analytics.
5. How masking works
Each detected item gets a label, and the same person always gets the same label. The AI understands the meaning without seeing the data. When you paste the answer back, the tool swaps the labels for the real data.
John Smith (SSN 123-45-6789) signed the contract. We pay John Smith £1,200 to GB82 WEST 1234 5698 7654 32.
[PERSON_1] (SSN [SSN_1]) signed the contract. We pay [PERSON_1] £1,200 to [IBAN_1].
6. Accuracy — and its limits
- Rules + local model (extension and web tool): 98.2 % of personal data found in 264 test documents, 95.6 % in hard texts, rules alone 97.1 %.
- Real public forms and contracts (99, filled in with made-up data): 98.2 %.
- Blind test — 30 new documents the engine had never seen before the measurement: 92.9 % (rules alone 88.6 %). This is the most honest number for unknown texts.
- 0 false hits on 40 texts without personal data.
- The test documents were written by separate helpers who never saw the engine; the numbers are re-checked after every change. Results on your texts may differ.
- So we promise lower risk, not perfect protection: doubtful words are shown in a “Check” row, and you should review the highlighted parts before sending and add anything that is missing (or put it on your known-names list).
7. The rest of the site
- No cookies, analytics or tracking. The site is hosted on Cloudflare, which like any server keeps technical logs (e.g. IP address) for security.
- AI Academy: exam answers are sent to our server for grading, because the correct answers are not in the browser. Your name is never sent — the server signs only its fingerprint (SHA-256). The server stores nothing about you; your progress is in your browser. It only keeps the total number of certificates issued (shown on the Academy page) — without IP address, name or certificate number.
- Private tools (OCR, PDF, speech, local chat …) run in your browser; models are downloaded once and stored locally.
- Newsletter only if you subscribe yourself (Buttondown).
Found data the tool did not hide, or anything that doesn’t match what is written here? Write to us: hello@hypervision.si
Accuracy measured: 2026-10-03.