This page is for data protection officers and security teams assessing ChatMasqué before allowing it. It describes what the extension does, what it does not do, and how to check it.
In short
- Everything happens on the device. Detection, tokens and mappings stay in the browser. There is no ChatMasqué server, no account, no telemetry.
- A single network connection: downloading the name-detection model from Hugging Face, at a pinned revision, verified file by file.
- Encrypted mappings (256-bit AES-GCM, non-exportable key), erased automatically after the chosen period.
- AI sites only see the masked text. The real values shown in replies remain unreadable to them.
How a piece of data travels
- Paste or drop. On a monitored AI site, the extension intercepts the paste (or the dropped file) before the page receives it. If the analysis fails, nothing is pasted in clear text without an explicit action by the user.
- Local analysis. The text goes to the extension’s engine, which runs in an isolated extension document (an offscreen document in Chrome and Edge, a frame of the background page in Firefox), never in the site’s page. Rules (emails, IBANs, identification numbers from several countries, technical secrets…) and a name-recognition model, run locally, find the personal data.
- Tokens. Each value is replaced by a token (
[PERSONNE_1],[IBAN_1]…). The token → value mapping is encrypted, then stored in the extension’s IndexedDB database. - Sending. The site receives the masked text, and only that.
- Reply. Tokens in the reply are replaced on screen by the real values, inside a closed shadow DOM: the site’s script cannot read them.
- Erasure. Mappings expire after 1, 7, 30 or 90 days (7 by default). A workspace can be purged at the end of an assignment, and everything can be erased in one click.
The pasted text itself is never stored. The optional history only keeps the date, the site and the number of masked values per type.
Images
ChatMasqué does not read what an image shows. By default, an image pasted or dropped on an AI site therefore goes through an editor where the user blacks out by hand what must not be sent. Blacked-out pixels are replaced, not blurred. The image is sent under a neutral name and, for JPEG, PNG or WebP, without its metadata (GPS location, device, date taken). An organization can also allow images without asking, or refuse them.
Network connections
- Name-detection model: downloaded once from the Hugging Face Hub (
huggingface.coand its*.hf.coservers), at a pinned revision. Each file is checked against its SHA-256 fingerprint: a modified file is rejected. - Nothing else. The content security policy (CSP) of the extension’s pages restricts connections to these addresses only. All the code ships in the published package: no code is loaded remotely.
Isolation from sites
- Elements added to pages (buttons, restored values, dialogs) live in closed shadow DOMs. Only a real user click counts: a page script can neither trigger “Paste without masking” nor read a real value.
- Every message received from a page is validated, and its origin compared with that of the sending tab. The extension exposes no resource and no entry point to sites.
- The encryption key, the mappings and the model are never reachable from scripts injected into pages.
Permissions
Each permission the extension requests is explained in the privacy policy. Known AI sites are recognized out of the box. Another site can only be added over HTTPS, at the user’s request, one site at a time.
For organizations
- Enforced configuration through policy (GPO, Intune, Google Admin, Edge): settings, data types, organization-specific entities, sites. Whatever is enforced is locked in the panel.
- Status attestation exportable for audits: version, protection, sites, types, SHA-256 fingerprint of the enforced configuration. It contains no real data.
- Encrypted backup with a passphrase (PBKDF2-SHA-256, 600,000 iterations, then AES-GCM). The file is unreadable without it.
- Licence: use within an organisation, like any professional use, requires a licence (see the terms of use); adapted terms for non-profits, education and the public sector.
Controls and tests
The reference framework is Google’s security guide for extensions (“Stay secure”). Each control is mapped to an OWASP Top 10 category and a CWE weakness. Fourteen controls are verified by automated tests for every release: minimal permissions, CSP, no remote code, no innerHTML or eval, message validation, resistance to hostile pages, encryption at rest, backups. There are three runtime dependencies, monitored weekly. A software bill of materials (CycloneDX SBOM) is produced for every release and available on request.
Limits
ChatMasqué greatly reduces the risk; it does not remove it.
- Detection is not infallible. A rare name or an unusual identifier can slip through. The preview before pasting lets you check, and learning mode asks the AI to flag what was left in clear text.
- Text typed on the keyboard is not masked automatically: the extension flags the data and offers to mask it.
- Context alone can identify a person: a rare job title, a small town and a date, even without a name. Masking does not replace thinking about what you send.
- What images show is not read: only areas blacked out by hand disappear.
- An already compromised device (spyware, malicious extension) is not protected by ChatMasqué.
Reporting a vulnerability
Write to developer@chatmasque.com, subject “ChatMasqué – security”, without opening a public ticket and without attaching real personal data. Acknowledgement within 3 business days, first assessment within 10 business days, fix within 30 days for a serious flaw.