Google tests broader permissions for Gemini on the desktop

Google is working on giving its AI chatbot Gemini broader permissions when it runs as an app on the computer. This could mean that Gemini will be able to carry out longer tasks across multiple programs without the user having to approve each individual step.

The discovery comes from TestingCatalog, which found a new setting called "Additional sandbox options" in a test version of Gemini desktop. A sandbox is a confined area where a program can work safely without harming the rest of the computer. The setting suggests that Gemini will be able to work outside selected folders, go online and use other programs without asking for permission every time.

Important actions still require approval

Even though Gemini gets more freedom, the user will still have to approve sensitive actions such as purchases or signing legal documents. The feature is only in testing right now, and Google has not said when regular users can get it.

The timing is interesting because Google recently launched a new AI model called Gemini 4 Argon. It is initially being tested by selected security experts through Google's Fairwind program. Google has not said when Gemini Ultra subscribers will get access to the new model.

Argon is also being tested specifically on controlling a computer autonomously, including through tests called Agent's Last Exam and OSWorld 2.0. That makes the model a possible candidate to power the more autonomous features in Gemini desktop. However, Google has not confirmed that Argon and the new features in the desktop app are connected.

Part of a bigger plan

The development fits into Google's overall plan for Gemini on the computer. Gemini Spark is already intended as a personal assistant that can research, plan and work with files and Google services over multiple steps. Broader sandbox permissions could turn the desktop app into a far more autonomous AI assistant that can work across the entire computer while still asking for permission on important decisions.

For now, the features are still in development, and their presence in a test version gives no definite answer as to when they will be available to everyone. Some Gemini features move quickly from testing to launch, while others can be held back for months.