Best AI Privacy Extensions for Chrome: 6 Tested on ChatGPT, Claude & Gemini
Six extensions that mask names, emails, ID numbers and other sensitive data before an AI chat sees them. We installed each one, pasted the same text in five languages into ChatGPT and recorded what it caught, what it got wrong and how it handled the AI's answer.
Below: the results, the features that matter for ChatGPT, Claude and Gemini privacy at work, and the prices.
Tested on 15 November 2026 by Andrii Kozak; listings checked 14 November 2026; updated 17 November 2026. We make ChatDataMask. All six extensions were tested on their free plans with default settings. Anything we did not test comes from the vendor's own pages, and features in beta or marked "coming soon" count as not available. Extensions change often, so check the current version before you install.
Which one to pick
Based on our test and on what each vendor lists.
- For a complete free workflow in English, German, Polish, Spanish or Ukrainian: ChatDataMask. In our test it found 99% of common personal data with no traps masked and no AI model, and it was the only one to cover every step, from typing to the restored answer. The only one that lists Polish and Ukrainian ID formats, on its Pro plan.
- For PDFs, scans and Office files: Privacy Mask. The only one here that lists OCR and redaction of Word, Excel and images; works in 8 AI chats.
- For a company-wide rollout: PiiBlocker. The only one here that lists managed deployment and an admin dashboard; works in 7 AI chats.
- For French, Italian or Dutch speakers: ChatWall. Interface in 6 languages, the most here; you can type in the chat or in a separate, isolated window. IBAN, card numbers and government-issued numbers need a paid plan.
- For the most detection, if your computer can run an AI model: Privacy Guardrail. The only one here that runs a large AI model by default; in our test it found the most government-issued numbers, but is a 160 MB install, checks pasted text only and restores answers in parts. Open-source public beta from DFKI, the German Research Center for AI.
- For a one-time payment instead of a subscription: Caviard. $9 per device for unlimited use on ChatGPT and DeepSeek. The free plan protects only 10 values per document, while each one-page text in our test had 16–18.
Test results: what each extension caught
Same text, same chat, free plans, default settings. Every column covers all five language sections (US English, German, Polish, Spanish, Ukrainian) together, except the second, which shows the US English section on its own. Higher is better for the first three columns, lower is better for the last one.
Swipe the table sideways to see every column →
| Extension | Common personal data65 values, 5 languages | Common personal data13 values, English (US) text | Government-issued numbers22 values, 5 languages | Traps masked by mistake14 traps, 5 languages |
|---|---|---|---|---|
| ChatDataMask | 99% | 100% | 5%A Pro feature; we tested the free plan | 0 of 14 |
| Privacy GuardrailAI model always on | 96% | 88% | 95% | 3 of 14 |
| PiiBlocker | 55% | 54% | 32% | 1 of 14 |
| Caviard | 52% | 54% | 9% | 0 of 14 |
| ChatWall | 49% | 46% | 50% | 2 of 14 |
| Privacy MaskOptional AI model, off by default | 41% | 54% | 36% | 0 of 14 |
Common personal data: names, companies, street addresses, emails, phone numbers, IBANs, card numbers, money amounts and dates of birth.
Government-issued numbers: national ID, tax, VAT and social security numbers, company registry numbers and licence plates.
Traps: reference and invoice numbers, and ordinary words such as "rose" in "the fee rose slightly", that should be left alone.
Percentages count a value as found when it was fully masked, and as half when only part of it was hidden. How we scored, and the limits of this test →
- ChatDataMask missed only part of one value, a Spanish industrial-estate address, and masked none of the traps. Outside the traps it did mask two ordinary Polish words as if they were names.
- Privacy Guardrail, the only one running a large AI model by default, found 96% of common personal data and the most government-issued numbers (95%), but masked 3 traps and marked job titles and other ordinary words as personal data more often.
- The other four found 46–54% of common personal data on the English (US) text, the language all six support, and 41–55% across all five languages.
Read these numbers as a snapshot, not a guarantee. Each language section is a short business text with about 15–20 values, run once. On longer documents or other kinds of text, results can differ from ours. Check the masked text before you send it.

Want to check your own text the same way? Add ChatDataMask to Chrome. It is free, with no usage limits.
Features compared: what we tested and what vendors list
Swipe the table sideways to see all six extensions →
| Feature | ChatDataMask | Privacy Guardrail | PiiBlocker | Caviard | ChatWall | Privacy Mask |
|---|---|---|---|---|---|---|
| Checked in our test, 15 November 2026 | ||||||
| Detects while you type | YesAt each line break, and on paste | NoPaste only | YesIn real time | YesIn real time | YesIn its own window; a warning sign in the chat's field | In Auto modeWhen you stop typing; Auto is off by default |
| Where you type | In the chat's field | In the chat's field | In the chat's field | In the chat's field | EitherIn the chat's field, or in its own window, opened before typing or on Send | In the chat's field |
| Mask or unmask one value | YesMask any word by hand; undo with a click or Ctrl+Z | On paste onlyCannot undo after inserting | Yes | Through settingsMasks automatically; edit the list to change it | Yes | All at onceSingle values could not be masked or undone separately |
| Warns before you send | Stops SendAsks before unmasked data goes out | NoChecks on paste only | Stops Send | No | IndicatorWarning sign next to the field | IndicatorWarning sign next to the field |
| Keeps paragraphs after masking | Yes | Yes | NoParagraphs merged | Yes | NoParagraphs merged | Yes |
| Restores the AI's answer | Whole answerOne value at a time, or all when copying | In partsSeparate reveal panels in one answer | NoThe answer stays masked | NoThe answer stays masked | Whole answerUnmask and copy the last response, or one tag at a time | Whole answerHover a value, or copy the latest answer |
| Copies the restored answer | YesPaste into Word or anywhere | In partsPanel by panel | No | No | YesWith its Unmask & Copy option | Yes |
| Masks are remembered | Until you clear themSurvived a browser restart | YesSurvived a browser restart | Not applicableThe answer stays masked | Not applicableThe answer stays masked | YesSurvived a browser restart | YesSurvived a browser restart |
| From each vendor's listing and website, 14 November 2026 | ||||||
| Feature | ChatDataMask | Privacy Guardrail | PiiBlocker | Caviard | ChatWall | Privacy Mask |
| Price | FreePro $4.99/month | FreeOpen source | FreeEnterprise from $6.99/seat/month, annual | Free$9 one-time per device | FreeSolo $4.90/month; teams from $19/month | FreePro $9/month or $60/year |
| Free plan | No usage limitsGovernment-issued numbers on Pro | FreeOpen source; no paid tiers listed | No usage limits | Limited10 values per document; ChatGPT and DeepSeek only | Basic types onlyNames, contacts, places, dates; IBAN, cards, IDs and keys paid | Limited200 hides and 10 files per week |
| AI chats supported in place | 3ChatGPT, Claude, Gemini | 3ChatGPT, Claude, Gemini | 7ChatGPT, Claude, Gemini, Grok, DeepSeek, Kimi, Perplexity | 4ChatGPT, Claude, Gemini, DeepSeek | 6ChatGPT, Claude, Gemini, Copilot, Grok, DeepSeek | 8ChatGPT, Claude, Gemini, Copilot, Google AI Mode, Grok, DeepSeek, Perplexity |
| Standalone editor for any text | YesSide panel, full-screen editor and Restore tab | No | No | No | NoIts window works over AI chats | Side panel onlyWorks with the open AI chat |
| Government-issued numbers listed | 5 countriesUS, Spain, Germany, Poland, Ukraine | Not listed | 2 countriesSSN, UK National Insurance | Not listed | Paid plansID, passport, SSN, VIN, plates; countries not stated | 9 countriesUS, UK, Spain, Italy, France, Germany, Netherlands, Russia, UAE |
| PDF, Word, scans | NoWarns about attachments | No | Team plansText-based PDFs; no OCR | Not stated | Not stated | YesPDF, Word, Excel, images; OCR in English, Spanish, Russian |
| Mask your own values | YesMark in the text; remembered automatically | YesAllow and block lists | YesRight-click; personal dictionary | YesMark in the text | YesRemembering them for next time is paid | YesList of custom terms |
| Interface languages | 5English, German, Polish, Spanish, Ukrainian | 1English | 1English | 1English | 6English, German, Dutch, Spanish, French, Italian | 3English, Spanish, Russian |
| Detection engine | Rules with validationNo AI model | RulesPlus an AI model, always on | RulesPlus an NER model | Pattern recognition | Rules and language processingOptional AI model, 0.8–2.7 GB | Rules with checksumsOptional AI model, about 1 GB |
| Install size | 1 MB | 160 MB16 GB RAM and WebGPU recommended | 1.6 MB | 0.2 MB | 9.8 MB | 25 MB |
ChatGPT, Gemini and Claude together had 93% of web visits among the seven largest AI chatbots in October 2026 (Similarweb data via Momentic). Sources: ChatDataMask · Privacy Guardrail · PiiBlocker · Caviard · ChatWall · Privacy Mask on the Chrome Web Store, plus the vendors' own pages: Privacy Guardrail on GitHub, PiiBlocker technical overview, Caviard pricing, ChatWall pricing, Privacy Mask. Listings checked 14 November 2026; the rows under "Checked in our test" were tested on 15 November 2026.
Each extension in brief
ChatDataMask Our product
Built for professionals and teams in Europe who handle client data: recruiters, lawyers, freelancers, analysts.
In our test: found 99% of common personal data on the free plan, with no traps masked. Government-issued numbers are a Pro feature and were not tested.


Strengths
- 35+ data types; country packs for the US, Spain, Germany, Poland and Ukraine
- Detects as you type, stops Send for unmasked data, restores the whole AI answer
- Side panel with a full-screen editor and a Restore tab, for text going anywhere
- Lightweight: no AI model download; masks are kept until you clear them
Limitations
- Does not scan attached files yet
- Works in place on 3 AI chats; others through the side panel
Privacy Guardrail
A research project from DFKI with open-source code.
In our test: found 96% of common personal data and the most government-issued numbers (95%), but masked three reference numbers. Checks were noticeably slower, and restored answers appeared in several separate panels.
Strengths
- Open source and free
- Multilingual NER model
Limitations
- Public beta; reviews pasted text only, not typed text
- The local AI model recommends 16 GB RAM and WebGPU
- ChatGPT, Claude and Gemini only
PiiBlocker
A free masker that works in 7 AI chats, with a path to managed deployment.
In our test: found 55% of common personal data and masked one trap, the Spanish month "abril", which fits a tool that supports English text only. Masking merged paragraphs, and the AI's answer stayed masked.


Strengths
- Works inside 7 AI chats
- Free without limits; enterprise plan with an admin dashboard
- Confirmation dialog before sending
Limitations
- Masks expire after 4 hours or when the browser closes
- No EU government-issued numbers listed
- PDF redaction only on team plans
Caviard
A simple, lightweight masker.
In our test: found 52% of common personal data, masking automatically without asking. It does not restore the AI's answer; the mapping is visible only in its settings.


Strengths
- Keyboard shortcut to switch between masked and original text
- One-time $9 licence instead of a subscription
Limitations
- Free plan limited to 10 protected values per document
- Free plan works on ChatGPT and DeepSeek only
ChatWall
A Luxembourg-based anonymiser with a focus on GDPR and digital sovereignty.
In our test: found 49% of common personal data. Its Unmask & Copy option restored the whole answer, both on screen and in the copy.
Strengths
- Optional isolated window that the chat site cannot read
- Interface in 6 languages; source code available on GitHub
- Detects as you type, with a warning sign next to the field
Limitations
- IBAN, card numbers, government-issued numbers and API keys need a paid plan
- Optional local AI model, 0.8–2.7 GB; the extension itself is about 10 MB
Privacy Mask
The most complete document workflow in this group.
In our test: found 41% of common personal data with default settings (local AI model off) and masked no traps. Selecting single values still masked the whole text.


Strengths
- PDF, Word, Excel, CSV and images, with OCR for scans
- 8 AI chats and a built-in text editor
- Custom terms and national IDs for 9 countries
Limitations
- Free plan limited to 200 hides and 10 files per week
- Pro costs $9 per month
- No Polish or Ukrainian identifiers listed
How we tested
- Who tested. Andrii Kozak, who builds ChatDataMask and has 12+ years in software testing, installed and ran every extension himself in Chrome. Each run was scored by the same script against a fixed answer key, so every extension, ours included, was judged by the same rules.
- Test text. Five sections, in US English, German, Polish, Spanish and Ukrainian, built from synthetic invoices, contracts, CVs and bank letters. 87 values had to be masked; the IDs, IBANs and card numbers among them pass their official checksums.
- Setup. One extension enabled at a time, free plan, default settings, a new ChatGPT chat for each run. Privacy Guardrail's AI model cannot be switched off, so it ran with it. Caviard was tested in two halves because its free plan stops at 10 values.
- Scoring. Each value counts as found, half found or missed. For addresses, the street and number must be hidden; city, postal code and floor are optional. Labels such as "EIN", "DNI", "CIF", "PESEL", "NIP" or "ІПН", and legal forms such as "LLC", "Corp.", "GmbH", "S.A.", "Sp. z o.o." or "ТОВ", are not counted. Masking a trap counts as an error; masking a city or an invalid card number does not.
- Limits. The test covers the five markets ChatDataMask is built for, which gives it a home advantage; it does not cover UK, Italian, French or Dutch identifiers. Paid plans and file redaction were not tested, and each text was run once, so a single value can move a result by several points.
Test texts and raw results
Run the same test on any extension, or check our scoring value by value. The texts are synthetic: every name, company, address and number is invented.
- Test texts: the five language sections exactly as we pasted them into ChatGPT.
- Results by value (CSV): every value and trap, what was expected, and whether each extension masked it, partly masked it or missed it.
Two kinds of AI privacy extension
Tracker and ad blockers, such as Privacy Badger or uBlock Origin, stop websites from following you around the web. They do not look at what you type, so a client's name or ID number in your prompt still reaches the AI company. Masking extensions, compared on this page, check your text and replace personal data with tags like NAME_1, so you can anonymize prompts before you send them. For ChatGPT privacy at work, many people use both.
What to look for
- Where detection runs. A privacy tool should find personal data on your device, so your text is never sent to an outside server just to be checked. Look for local processing in the listing, and no account for the free version.
- Whether it catches what you type. Some tools check only pasted text. If you draft prompts by hand, you need detection while you type, before you press Send.
- Formats from your country. Every language writes names, addresses and phone numbers its own way, sometimes in another alphabet, and ID numbers differ from country to country. Check which languages and countries the tool lists, not only "ID numbers".
- Whether the answer still makes sense. Good tools give the same value the same mask and put the real values back into the AI's reply, so you can use it straight away.
- How long masks are remembered. If you come back to a chat tomorrow, can the tool still restore the answer? Some keep masks for hours, some until you clear them.
- What it costs your computer. Tools with a local AI model can need gigabytes of memory, take time to load and slow the browser down while they work. Rule-based detection is lighter, starts instantly and keeps typing smooth.
- Whether IT will approve it. On company laptops, large model downloads and GPU access are often restricted. A tool that needs neither is usually easier to get approved and to roll out.
When an extension is not enough
A browser extension protects what you type in Chrome. It does not cover the ChatGPT desktop or mobile apps, other browsers, or files you upload, so mask the text of a document before you attach it. If your company needs central policies, audit logs and coverage across every device, look at enterprise platforms such as Prompt Security (now part of SentinelOne), Nightfall AI or Microsoft Purview, usually alongside an extension rather than instead of one.
Looking for an alternative?
Privacy Guardrail alternative
Privacy Guardrail checks pasted text only, runs a 160 MB install with an AI model that recommends 16 GB of RAM, and restores answers in separate panels. ChatDataMask checks text as you type, needs no AI model and restores the whole answer in one go.
PiiBlocker alternative
In our test PiiBlocker left the AI's answer masked and merged paragraphs after masking, and its masks are cleared after 4 hours or when you close the browser. ChatDataMask restores the whole answer, keeps masks until you clear them and lists government-issued numbers for five countries, including Poland and Ukraine.
Caviard alternative
Caviard's free plan works on ChatGPT and DeepSeek, with up to 10 protected values per document, and it does not restore the AI's answer. ChatDataMask's free plan has no usage limits, works in ChatGPT, Claude and Gemini, and puts the real values back into the answer.
ChatWall alternative
ChatWall's free plan masks names, contact details, places and dates; IBAN, card numbers and API keys need a paid plan. ChatDataMask masks all of those for free, both in the chat's own message field and in a standalone side-panel editor for text going anywhere.
Privacy Mask alternative
Privacy Mask limits its free plan to 200 hides and 10 files per week, and Pro costs $9 per month. ChatDataMask has no usage limits on its free plan, and Pro costs $4.99 per month. If you mainly redact PDFs and scans, Privacy Mask remains the stronger choice.
Questions
Which Chrome extension is best for AI privacy?
It depends on your data. In our test, ChatDataMask's free plan found 99% of common personal data across five languages with no traps masked, without an AI model, and it is the only one that lists Polish and Ukrainian ID formats, on its Pro plan. Privacy Guardrail found the most government-issued numbers with its large AI model, if your computer can run it. For PDFs and scans, Privacy Mask is the only one with OCR.
Which extension found the most personal data in your test?
It depends on the type of data. On common personal data such as names, addresses, emails, phones, IBANs and cards, ChatDataMask's free plan found the most (99%) and masked none of the traps. On government-issued numbers, Privacy Guardrail, which runs a large AI model by default, found the most (95%); in ChatDataMask these are a Pro feature, which we did not test. The other four found about half.
Do these extensions put the real values back into the AI's answer?
In our test, ChatDataMask, ChatWall and Privacy Mask restored the whole answer and let you copy it with the real values. Privacy Guardrail restored it in separate panels, and PiiBlocker and Caviard left the answer masked.
Is Privacy Badger or uBlock Origin enough to protect what I type into ChatGPT?
No. Tracker and ad blockers stop websites from following you, but they do not look at the text you send to an AI chat. To keep names, emails or ID numbers out of your prompts, you need a masking extension like the ones compared here.
Are these extensions free?
All six have a free option. ChatDataMask, PiiBlocker and Privacy Guardrail have no usage limits on their free plans. Privacy Mask limits Free to 200 hides and 10 files per week, and Caviard to 10 protected values per document. ChatWall's free plan masks names, contact details, places and dates; IBAN, card numbers, government-issued numbers and API keys need a paid plan from $4.90 per month.
Do these extensions send my text to a server?
According to their Chrome Web Store listings, all six detect and mask personal data locally in the browser. Paid plans may use servers for sign-in and billing, not for scanning your text.
Can they mask PDFs, Word files or scans?
Privacy Mask redacts PDF, Word, Excel and images, with OCR for scans. PiiBlocker redacts text-based PDFs on its team plans. ChatDataMask works with text and warns you when a message has attachments, so you know what goes out unchecked.
Which extension supports Polish, Ukrainian, German or Spanish ID numbers?
ChatDataMask Pro detects PESEL, NIP, REGON and NRB for Poland and RNOKPP (tax number), EDRPOU and UNZR for Ukraine. None of the other five list Polish or Ukrainian identifiers, although in our free-plan test Privacy Guardrail's AI model caught them and ChatWall caught some. ChatDataMask's free plan, with no AI model, caught nearly all names, addresses, IBANs and other common data in all five languages it lists. For Germany, it also detects tax ID, social security, health insurance and driver's licence numbers, and for Spain DNI/NIE, NIF/CIF and social security numbers.
How did you compare these extensions?
We installed all six on 15 November 2026 and pasted the same text, in five languages, into ChatGPT with one extension at a time, on free plans and default settings. Features we did not test come from each extension's Chrome Web Store listing and website, checked 14 November 2026. We make ChatDataMask; the full method and its limits are under "How we tested".
Try ChatDataMask free
No limits on the Free plan. Works right inside ChatGPT, Claude and Gemini.



