Anthropic added a built-in Chromium-based browser to Claude's Cowork desktop app on Mac, Windows, and Linux, rolling out to Pro, Max, and Team subscribers. This reduces reliance on the Claude Chrome extension for many web tasks, though the extension remains useful for working with pages already open in a user's own browser and stays the default if installed. The new browser can import logins from Chrome, Edge, and Firefox (with OS-specific limitations), but explicitly excludes banking, email, and SSO logins. Anthropic acknowledges prompt injection risks remain nonzero despite built-in safeguards.
Questions this post answers
Does Claude's Cowork app now have its own built-in browser instead of relying on the Chrome extension?
Yes, Anthropic added a built-in Chromium-based browser to Claude's Cowork desktop app on Mac, Windows, and Linux, rolling out to Pro, Max, and Team plan subscribers. Previously Claude needed the Claude in Chrome extension to browse the web; now it has its own separate browser for tasks that don't require your existing browser session. The extension remains the default if already installed, and is still preferred for working with pages you already have open. Track how Claude's tooling evolves for agentic browsing workflows on daily.dev.
Can Claude's built-in browser in Cowork access my saved logins from Chrome or Firefox?
Yes, Anthropic lets Claude's built-in browser import logins from Chrome, Edge, and Firefox on macOS, and from Firefox on Windows and Linux, but banking, email, and single sign-on logins are explicitly excluded. The inconsistency across operating systems stems from Chrome and Edge on Windows blocking third-party apps from reading local cookies as an infostealer countermeasure, while Firefox stores them in a plain SQLite database and macOS Chrome keys sit in Keychain. Developers evaluating AI agent security tradeoffs can follow browser-permission changes like this on daily.dev.
What security risks come with letting an AI agent like Claude control its own browser?
Prompt injection remains a real risk whenever an agent browses the web autonomously, and Anthropic states plainly that despite built-in safeguards in both the Chrome extension and the new built-in browser, the risk is not zero. Because the built-in browser typically isn't logged into sensitive accounts like banking or email, the potential blast radius from a successful attack is expected to be smaller. Anyone weighing AI agent risk versus convenience can keep tabs on incidents like this via daily.dev.