The AI assistant most widely installed on desktop computers might be one step away from permission to do literally anything on your machine: read, create, delete files, browse the web unsupervised and control other applications. That is the implication of a clue found inside the Gemini desktop app's own code, which started circulating this week after researchers at TestingCatalog combed through the app looking for strings and functions that are not yet active.
How the clue surfaced
The discovery did not come from a Google announcement, but from reverse engineering: TestingCatalog found references hidden inside the Gemini desktop binary pointing to a broader permission mode, internally described as something close to full computer access. BleepingComputer replicated the analysis and reached the same conclusion, which lifts the clue from an isolated rumor to a more consistent signal: this is not just one source claiming to have seen something, it is a code artifact that anyone with the installed app can try to reproduce.
What would change in practice
Today, when Gemini on desktop wants to do something on the computer, it usually goes through one off confirmations: opening a specific file, running an action inside an authorized app. The mode found in the code suggests something closer to the computer use capability Anthropic built into Claude and that OpenAI is testing with agents: the assistant would get broad permission to browse the file system, create and delete documents, open the browser and operate other programs without asking for authorization at every step. For people using Gemini to automate repetitive tasks, that would cut friction. For people worried about security, it is the classic recipe for an agent with too much privilege and no visible guardrail.
Between code and finished product
Having a hidden function inside the binary is not the same as having a feature ready for users. Unreleased code can be tested internally, dropped, limited to a closed waitlist or shelved for months, as has already happened with other Gemini features found before their official launch. Google has not confirmed the mode exists, has not detailed what kind of sandbox, user approval or permission scope would come with it, and there is no rollout date. The classification here is strong signal, not confirmation: there is evidence of real code, and two outlets reached the same reading, but the official word from Google about what the function does and when it leaves the lab is still missing.
MaxAssistant's take
The move fits where the whole industry is heading: models moving from answering questions to operating entire computers. After watching Anthropic and OpenAI push ahead with agents that control the screen and the system, it would be odd if Google were not trying something similar with Gemini, which already has the largest installed base of any desktop assistant. If this mode moves from code to product, the real differentiator will not be the permission itself but how Google designs the confirmation and rollback layer: an agent that deletes the wrong file without asking is a recipe for a bad headline, not for mass adoption.
Sources
TestingCatalog: Gemini desktop to get broader computer use permissions; BleepingComputer: Google Gemini could soon get full access to your Macs files, apps and the web.