Google gives Gemini computer-use capability

Google integrated computer-use capability directly into Gemini 3.5 Flash. Previously this was a separate Gemini 2.5 computer-use model; now the capability is part of the main Flash model. In practice, developers can build agents that work in browser, mobile and desktop environments, not only through programming interfaces. This is an important shift for enterprise deployment. Many corporate processes still run in applications without convenient automation interfaces. Employees click through internal systems, enter data into forms, check documentation, test interfaces and copy information between tools. A computer-use agent can take over such tasks where traditional integration is missing or would be expensive. Google lists examples such as longer-running enterprise automation, continuous software testing and knowledge work in professional applications. Developers can use the capability through the Gemini API and Gemini Enterprise Agent Platform. This suggests Google does not want to keep the function as a demo, but as part of an enterprise agent platform. The risk is obvious: an agent that can see and click can also cause damage. Google therefore emphasizes protections against indirect prompt manipulation. It mentions targeted adversarial training and two optional enterprise safeguards. The first requires explicit user confirmation for sensitive or irreversible actions. The second automatically stops a task if the system detects an indirect malicious instruction. The business impact will be largest for companies with many legacy applications and manual checks. Instead of long integration projects, they can try agents that work in existing interfaces. Legal accountability, however, does not disappear. Terms: Computer use is AI’s ability to control applications through screens, clicks and inputs. An indirect malicious instruction is a hidden instruction inside content. A sandbox protects production systems from agent errors.

Google gives Gemini computer-use capability

The full article is available to users registered at Hard Skills