Skip to content
AgentIDDocs

Type at least two characters to search titles, headings and content.

↑ ↓ to navigate↵ to open
AgentID Docs

Changelog

2026-10-14 release

Restoration across whole conversations, one Claude Desktop integration for Windows and macOS, a Masked values view in the activity log, and broader detection of personal data.

Highlights

  • Your values back, safely. The browser extension now restores masked values across the whole conversation, including earlier messages and after a reload. The AI provider and AgentID's logs keep the placeholders.
  • One Claude Desktop integration. A single Claude Desktop target protects Chat, Cowork and Code on Windows and macOS, and Claude Code in the terminal.
  • See what the AI received. The activity log shows a Masked values table for browser turns, with originals stored encrypted and revealed only through an audited action.
  • Better detection of names, addresses, account numbers and Czech and Slovak business identifiers, and of long documents.

Dashboard

  • The + Register New AI System wizard can show download buttons and the first setup steps for the browser extension and Claude, where AgentID hosts the client packages.
  • New Claude Desktop target with a simpler setup: Download Claude plugin (already connected to your system), a one-line install command per computer, and a Whole organization option with an organization plugin and Windows and macOS deployment packages.
  • Activity (Logs) shows only real prompts and counts them accurately, gives each prompt row one consistent status and category, and shows any prompt with a logged answer as Complete. The extension's Test connection is logged as a health check instead of a prompt.
  • Masked values in the log detail: the token, what the AI provider received, and the original (encrypted, shown after Reveal payload). Originals are stored only with Store Encrypted Payloads on.

Masking and detection

  • Typed placeholders only. Masked values always reach the AI provider as placeholders such as <PERSON_NAME_1>. Guard reports client_substitute_style: "tokens".
  • All contact details are masked, including work e-mail addresses and phone numbers.
  • Person names: initials are masked, courtesy titles stay outside the name, split names are joined, and ordinary words are no longer taken for names.
  • Addresses end where ordinary prose begins; account numbers introduced by a clear cue are masked.
  • Czech and Slovak documents: a labelled IČO keeps its type (also with a failed checksum) and the address behind it stays separate, a DIČ is one VAT_ID, health insurance numbers are HEALTHCARE_ID, and birth-year phrases such as "ročník 1975" or "rok nar." are masked as YEAR_OF_BIRTH.
  • Long documents: detection is robust to formatting; texts of 50,000 characters or more use the deterministic detection layer; fail-open systems degrade step by step instead of sending text unmasked, while fail-closed systems still block on a detection failure.
  • Secrets that cross a chunk boundary of a long text are masked whole.

Guard API

  • New optional pii_detection_mode (full or minimal). minimal is honoured only as one fast retry of a slow full check of the same request within 60 seconds. See Long texts and documents.
  • New response field client_substitute_style (always tokens) when restoration is enabled.
  • The OpenAPI description now documents prompt_context, attachments, origin, metadata, transformed_attachments, input_char_count and transformed_input_complete, and the correct input limit of 64,000 characters.

Browser extension 0.1.11

  • Restores masked values across the conversation with a local, encrypted store, also after reloading a Copilot conversation.
  • Answers to pasted text and files in Claude are captured and logged.
  • Replies are captured in background tabs and in more ChatGPT layouts.
  • A long masked paste is ready to send immediately; a checked ChatGPT prompt is sent without a second press.
  • A slow Guard check is retried once in the fast minimal mode; the extension never sends an unmasked prompt because of a slow check.
  • Users who share one AI system no longer cancel each other's checked prompts.

Claude

  • AgentID Security plugin 0.9.6 for Cowork and Claude Code: connected one-click download, organization-wide rollout as a required plugin, Node.js found outside the app's PATH (and a visible notice when it is missing), one fast retry of a timed-out check, and answers that stay visible.
  • Windows companion 0.5.4 for Claude Desktop chat: one PowerShell install command, a status line above Claude's message box, and logging of long multi-step tasks.
  • macOS companion 0.2.1 (new): a native AgentID Claude app installed with one Terminal command; it reliably verifies the protected prompt before sending it.