Upcoming versions of Google's Gemini may gain the ability to access files anywhere on a macOS system, launch applications, navigate the web, and execute tasks with fewer interruptions for user approval. According to findings by TestingCatalog on X, Google is currently experimenting with desktop control functionality for Gemini, including references to a previously undisclosed "Additional sandbox options" setting within the Gemini Desktop application.
The capability remains in testing and has not been officially announced by Google, but once activated, Gemini could read, create, modify, or delete files throughout your machine. This would extend Gemini's reach beyond directories that users have deliberately shared with the application.
The rollout trajectory remains uncertain, particularly given Apple's efforts to restrict how AI agents interact with personal files and information stored on macOS devices.
The Gemini Desktop interface indicates that Google's AI would also be capable of interfacing with applications including Mail, Safari, and Messages to carry out operations through those programs. Google states in the undisclosed settings panel: "By enabling additional sandbox options, you will be able to expand what Gemini can do and access on your Mac." The company further notes that "Depending on which settings you enable, Gemini may be permitted to take actions without asking for your permission first."
Gemini would still ask before sensitive actions
Google is not planning to grant Gemini unrestricted system control without protections in place. The experience would resemble Claude's approach, where users explicitly authorize the AI to interact with their computers.
Gemini would continue to request user confirmation before undertaking sensitive operations such as purchasing items, moving funds, establishing new accounts, consenting to legal agreements on the user's behalf, or altering personal information.
This feature appears to align with Google's larger vision for computer-use capabilities in Gemini, enabling the AI to operate across documents, web services, and installed applications rather than remaining confined to a conversational interface.
The timeline for launching Full Access and which Gemini model version will support it have not been disclosed.