> The most consequential new feature is Auto Browse, powered by Gemini 3, which handles multi-step tasks autonomously: scheduling appointments, filling forms, collecting documents, filing expense reports, and managing subscriptions across websites without requiring the user to navigate each step manually.
What could _possibly_ go wrong?
> Google built a double-check safety system that independently reviews the AI’s actions before executing them, with strict boundaries limiting the agent’s access to specific relevant websites and explicit user confirmation required for sensitive actions such as purchases or social media posts.
Like how they double check Workspace accounts, auto-logging them out periodically and returning the user to the login screen saying "prove it's you!", with both the login name and password fields pre-populated? _That_ kind of double-check?
Hasn't Windows' "do you really want to run this .exe file?" dialog long-since taught the world that confirmation fatigue is more a vulnerability than security mechanism?
