Security: what stays on your device

In short

• Your documents are processed in your browser. They are never sent to a server, including for the AI features.
• Nothing leaves your device during processing. You can check this with your browser's Network panel or by going offline.
• The restore key is encrypted on your device and stays there.
• The tool assists you. Review is mandatory, and the result remains your responsibility.

How the tool works

When you open Localex, your browser downloads the page and its code from the site's own domain. From then on, everything happens on your device:

1. reading the file and extracting the text;
2. detecting the data to mask;
3. replacing it with the strategies you chose;
4. review, leak scan and export.

The heavy work (reading PDFs, detection, text recognition, AI) runs in Web Workers: separate threads inside your browser. They pass their results only to the page.

Your documents stay in the tab's memory. Closing the tab erases them. In case-file mode (an unlocked feature), files can be kept in the browser's storage (IndexedDB). They are deleted automatically after 24 hours, and you can delete them sooner.

The AI features also run on your device. The model files are downloaded once, from the site's own domain, then kept in the browser's cache. No online AI is called and no content is sent to any API.

The site's Content Security Policy only allows the page to connect to the site's own domain. No third-party script is loaded: no analytics, no advertising, no widgets. You can read the policy in the page's response headers: Network panel, first request, Content-Security-Policy header.

What leaves your device, and when

Unlocking only happens if you ask for it. The confirmation e-mail is sent by Avlex. The data and retention periods are set out in the privacy notice.

Check it yourself

You do not have to take our word for it. Two tests are enough, with no special software.

With the browser's Network panel

1. Open the tool on the home page.
2. Open the developer tools: F12 or Ctrl+Shift+I (Windows), Cmd+Option+I (Mac). Select the Network panel.
3. Tick Preserve log, then clear the list.
4. Load a document, run the detection, review and export.

5. Look at the list:

   • every request goes to the site's own domain;
   • none contains your document, its text or the detected data;
   • if a request to /api/metrics appears, open it: its Payload contains counters only.

Offline test

1. Open the tool while online. First process a test document of the same format: some parts of the code are only loaded on first use.
2. Go offline: turn off Wi-Fi, unplug the cable or select Offline in the Network panel.
3. Load your document, process it and export it.

Processing succeeds without a network, because it depends on no server. Only unlocking and the first download of the AI files need a connection.

What the "0 bytes sent" meter measures

The meter adds up what the page sends while a document is being processed. It stays at 0 because processing sends nothing. The page computes it itself: it is an indicator, not proof. The two tests above do not depend on us.

Leak scan and hidden data

Before every export, the tool re-reads the file it produced, not the preview on screen. It searches it for every original value and its variants:

• DOCX: every part of the file, including headers, footers, footnotes, comments, document properties (docProps) and custom data (customXml);
• PDF: all text and the metadata.

The result is "0 leaks", or a list of occurrences that blocks the export until they are fixed.

The scan looks for the values the tool knows: those it detected and those you added. It cannot find data that was never spotted. That is why review remains mandatory.

The tool also cleans hidden data:

• DOCX: comments, tracked changes (accepted or rejected, as you choose), author and company metadata, thumbnails;
• PDF: Info and XMP metadata, annotations, attachments, form fields;
• images: EXIF data.

Images embedded in a document are flagged. You can extract their text with OCR (an unlocked feature), remove them, or keep them with a warning. Text inside an image is only scanned if it was extracted with OCR.

True redaction of PDFs

Fake redaction draws a black box over the text. The text is still underneath: it can be selected, copied or searched.

Localex never does this. It offers two PDF exports:

• Text PDF: the document is rebuilt from the pseudonymised text. The text stays selectable; the original values are no longer in it.
• Redacted PDF, original layout: each page is turned into an image, the masked areas are blacked out in the image, and no text layer is kept. The text is therefore no longer selectable, and the file is larger.

In both cases, the leak scan runs before the export.

The restore key

In pseudonymisation mode, the tool replaces each item with a token, for example [PERSON_1]. The table linking tokens to original values forms the restore key.

• Encryption: AES-GCM, 256-bit, through the browser's WebCrypto API.
• Passphrase: the encryption key is derived from your passphrase with PBKDF2-SHA-256, using 600,000 iterations.
• Storage: you save the key file to your disk. You can also keep it in the browser for the session. It is never sent.
• Restoring: on the Restore page, paste the AI's answer, load the key and enter the passphrase. The original values are put back on your device. Tokens the AI altered, for example [PERSON 1], are recognised and flagged.

If you lose the key or forget the passphrase, restoring is impossible. We have no copy and cannot rebuild it.

Keep the key separate from the pseudonymised text: that separation is what protects the people concerned.

Pseudonymisation or anonymisation

• Pseudonymisation: data is replaced with tokens, and a key lets you go back to the original. For you, as the holder of the key, the pseudonymised text remains personal data, and the GDPR still applies (GDPR, Article 4(5) and Recital 26; EDPB, Guidelines 01/2025 on Pseudonymisation, adopted on 16 January 2025 for public consultation).
• Anonymisation: no key is produced and the replacements are irreversible. Data is anonymous only if no one can reasonably identify the person any more, including by combining it with the context. A document can remain identifying without a single name: a rare profession, a small municipality, a well-known event.

The Court of Justice has clarified that identifiability is assessed according to the means reasonably available to each recipient (judgment of 4 September 2025, C-413/23 P). This changes nothing for you: you hold the key.

Limits

• No detection is exhaustive. Rules, lists and AI models can miss an item, or flag one wrongly.
• Human review is mandatory before any export. It cannot be switched off.
• The tool does not check the legal content of an AI answer.
• The security of your device, your browser and its extensions is outside the tool's reach. An extension can read the content of a page: use an up-to-date browser and keep extensions to a minimum.
• Processing is local, but the page and its code come from our server on every visit. The tests above let you check what they do.

The bars' AI guidelines

In January 2025, the Flemish Bar (Orde van Vlaamse Balies, OVB) and AVOCATS.BE (the French- and German-speaking bars, OBFG) published the same guidelines on lawyers' use of AI. In March 2026, the French-speaking Brussels Bar (OFABB) proposed a model AI charter for law firms.

These texts are addressed to you: you meet them, and the tool helps. The table shows, for each relevant requirement, what the tool does and what remains your responsibility. Localex is designed following these guidelines. Neither the OVB, AVOCATS.BE nor the OFABB has assessed or recommended the tool. The quotations below are our translations.

Sources:

• OVB, Richtlijnen voor advocaten voor het gebruik van Artificiële Intelligentie, 20 January 2025, Dutch and French text.
• AVOCATS.BE, Lignes directrices à l'intention des avocat(e)s sur l'utilisation de l'intelligence artificielle, published on 31 January 2025.
• OFABB, Charte d'utilisation responsable des outils d'intelligence artificielle, proposed model, March 2026.