Tech giant Google has announced that it is integrating its computer use tool directly into the core of its new Gemini 3.5 Flash model. This will make it easier to create autonomous AI agents that can interact with browser, mobile, and desktop interfaces without the need for third-party tooling.
Tech giant Google has announced that it is integrating its computer use tool directly into the core of its new Gemini 3.5 Flash model. This will make it easier to create autonomous AI agents that can interact with browser, mobile, and desktop interfaces without the need for third-party tooling.
The company announced the new feature on its official blog. With this update, developers can now use Gemini 3.5 Flash to create AI assistants that can independently analyze the screen, make decisions, and perform actions in real-world user environments.
Developers and enterprise customers can already test the tool via the Gemini API and Gemini Enterprise Agent Platform. A demo version of the Browserbase-based environment is also available.
Google notes that the update is focused on automating routine and long-term tasks. The model is able to analyze applications, classify their functions, and perform scenarios such as continuous software testing or collecting and processing information in professional software.
The developers paid special attention to cybersecurity, as giving AI models direct access to control the computer creates new vectors for hacker attacks. To counter threats, Google conducted targeted adversarial training of the model to minimize the risks of unauthorized input of hints.
In addition, the company is implementing two additional protection systems for corporate clients:
mandatory confirmation of actions by the user before performing sensitive or irreversible operations;
automatic stopping of task execution in case of detection of signs of indirect intervention of third-party instructions (indirect prompt injection).
Google experts also recommend that developers not rely solely on the model’s built-in security algorithms, but combine them with isolated virtual environments (sandboxes), human verification of actions, and strict access control.
Previously, the ability to control a computer was only available as a separate experimental model in the Gemini 2.5 line. The direct integration of this feature into the main and fast Flash-class model is part of the global trend in 2026 to move from simple chatbots to full-fledged AI agents capable of performing complex computer work in an autonomous mode.
