Google has launched Computer Use in the Gemini 3.5 Flash API, enabling the model to analyze on-screen content and autonomously perform actions in browsers, desktop operating systems, and mobile OSes. The model underwent adversarial training to guard against prompt injection attacks—when an exploit is detected, the task is automatically interrupted, and irreversible or critical actions require mandatory human confirmation. The tool is available through the Gemini API and the Enterprise Agent platform, with reference code and documentation published on GitHub alongside a demo environment hosted on Browserbase.

Introducing Computer Use in Gemini 3.5 Flash