2026-10-14 update

Highlights

The whole message is checked. When you paste long text into a message that already has text in it, the browser extension checks and masks the complete message, not only the pasted part.
Every text file is checked. Plain-text, CSV, JSON, e-mail, HTML, log and similar files, and PowerPoint presentations, are now checked and masked before upload.
Clear handling of files AgentID cannot read. For other file types the system's failure policy now decides: fail-closed systems do not upload them, fail-open systems upload them with a visible notice and a log entry.
Better detection of personal data in long documents, repeated values, multi-line and UK addresses, names in form fields, and Czech and Slovak identifiers.

Browser extension 0.1.12

Pasting long text into a non-empty message checks the whole resulting message. If you keep typing while the check runs, nothing is inserted, your typing is kept, and you are asked to paste again.
While a file is being checked, another selected, dropped or pasted file is not uploaded. You see a message and can add it again when the check finishes.
Text files (.txt, .csv, .tsv, .json, .xml, .html, .eml, .log, .yaml and other text/* files) and .pptx/.pptm presentations are checked like documents. The AI provider receives a masked copy named .txt.
Files AgentID cannot check (for example legacy Office, OpenDocument, iWork, archives, audio and video):
- Fail-closed systems: the file is not uploaded, and the message names the file type.
- Fail-open systems: the file is uploaded with a notice that it was not checked. The activity log records the file type and size, never the name or content.
When you select several files, nothing is dropped silently. The message lists what was uploaded, what was not, and why. Up to four files are checked at a time.

Masking and detection

Values repeated later in a document without their label, such as a birth date or ID, are masked with the same placeholder as the first mention.
Addresses split over two lines or quoted in a reply are masked as one address. UK addresses with a full postcode are recognised consistently.
Names given as the value of a field such as "Client" or "Klient", and more given names in Czech, Slovak and English, are detected.
Receipt and submission dates are no longer taken for birth dates, and the Czech word "žádost" is no longer masked as a name.
When the remote detection stage misses a clear local finding, such as a full postal address or a labelled birth date, the local finding is kept.
Very long texts (50,000 characters or more) are processed faster.
Long texts and documents are now also checked by the remote detection stage, section by section. Prompts wait at most about 5 seconds; documents and attachments use up to about 28 seconds. Sections that do not finish in time keep local detection.
Phone numbers, birth dates, addresses, birth numbers, IBANs, card numbers, UK number plates and company and VAT numbers are detected more reliably in Czech, Slovak and English, including tables, signatures and quoted replies.
Fewer false masks: amounts, invoice and order references and other dates stay visible.
Company and institution names are not masked, unless they contain a person's name. Variable symbols are not masked, unless they are a birth number. Policy, claim and customer numbers stay masked.

Security

Password reset and invitation e-mails: when an e-mail cannot be delivered, AgentID now reports a delivery error and asks you to try again. Links are only ever sent by e-mail.
Systems that require single sign-on for the browser extension now reject a request when the session's revocation status cannot be confirmed.
Single sign-on sessions are now also applied to documents uploaded for checking.
Dataset changes in the dashboard require administrator access to the AI system, and the audit log records only changes that actually happened.
Very old Claude Cowork plugins (before 0.7.6) are now checked like every other client. Update to plugin 0.9.6.