Claude Desktop seems to be a bit sneaky
-
While looking into another matter, the researcher discovered a Native Messaging host manifest on his Mac that he did not knowingly install. On Chrome and other Chromium-based browsers, extensions can exchange messages with native applications if they register a native messaging host that can communicate with the extension.
By testing on a clean machine, Hanff discovered that Installing Claude Desktop for macOS drops a Native Messaging host manifest into multiple Chromium profiles (Chrome, Edge, Brave, Arc, Vivaldi, Opera, Chromium), even including for browsers that are not actually installed yet.
The Native Messaging host manifest tells a Chromium‑based browser which local executable to invoke when an extension calls a native host, and those hosts run outside the browser sandbox with current users permissions. Hanff therefore describes this as a “backdoor.” The manifest pre‑authorizes three Chrome extension IDs, so any extension with those IDs can call the helper via connectNative, giving it access to browser automation features.
Another objection is that Claude makes simple deletion futile since the manifest will be recreated the next time the user launches Claude Desktop.
It’s important here to point out that his article is about Claude Desktop, the Electron-based macOS application with bundle identifier com.anthropic.claudefordesktop, distributed as Claude.app. It is not about Claude Code, Anthropic’s command line developer tool. Claude Code is autonomous (“agentic”), allowing you to hand over a task, and it handles the planning and execution until done. So, for Claude Code, it would absolutely make sense to enable communication with browsers, provided they are present on the target system.
So, we have an application that writes into other apps’ profile/support directories (the browsers’ configuration area) and can act as the user, with capabilities like using the logged‑in browser session, DOM inspection, data extraction, form filling, and session recording. This expands the attack surface of every machine this manifest is dropped on, without asking for consent.
Anthropic’s own launch blog on “Claude for Chrome,” which discusses Anthropic’s internal red‑team experiments, explicitly mentions prompt injection as a key risk and reports attack success rates of 23.6% (no mitigations) and 11.2% (with mitigations). Hanff cites this to argue that a pre‑positioned bridge is a non‑trivial risk.
Hello! It looks like you're interested in this conversation, but you don't have an account yet.
Getting fed up of having to scroll through the same posts each visit? When you register for an account, you'll always come back to exactly where you were before, and choose to be notified of new replies (either via email, or push notification). You'll also be able to save bookmarks and upvote posts to show your appreciation to other community members.
With your input, this post could be even better 💗
Register Login