Gemini 3.5 Flash: AI Agents in Computing Environments

Google has unveiled Gemini 3.5 Flash, a breakthrough in AI technology with the integration of ‘computer use’ capabilities, signaling a move towards autonomous agents capable of engaging with digital environments.

Enhancing AI with New Capabilities

Released on June 24, 2026, this update empowers developers to create intelligent agents that can operate across various platforms, including browsers, mobile devices, and desktop systems. Previously, such capabilities were confined to the standalone Gemini 2.5 model but now come built-in with Gemini 3.5 Flash.

This incorporation not only boosts performance but also streamlines the deployment process for developers aiming to automate tools and workflows in enterprise settings.

Advanced Features of Gemini 3.5 Flash

Gemini 3.5 Flash introduces sophisticated agentic behaviors, enabling AI systems to effectively ‘see, reason, and act’ within computing environments. These agents can undertake complex tasks such as navigating web interfaces, performing software tests, and managing enterprise applications.

The model extends the existing capabilities of Gemini, including function calling and integration with tools like Search and Maps. Notably, it achieved a score of 78.4% on the OSWorld-Verified scale (source: Google).

With its enhanced computer use capability, Gemini 3.5 Flash can now manage long-duration tasks requiring sustained interaction and informed decision-making.

Addressing Security Concerns

The introduction of these capabilities also enlarges the potential vulnerability landscape, especially in enterprise environments where AI agents have access to sensitive systems or workflows. In response, Google has implemented adversarial training to mitigate risks such as prompt injection attacks, which can alter AI behavior through malicious external inputs.

Additional security measures include mandatory user confirmation for sensitive actions and automatic task termination upon detection of indirect prompt injection attempts. Google advises a comprehensive defense strategy, combining these measures with sandboxing, strict access controls, and human oversight.

Despite these precautions, security specialists caution that autonomous agents interacting with live systems might present novel exploitation opportunities if not properly configured.

Gemini 3.5 Flash, equipped with computer use functionality, is accessible via the Gemini API and the Gemini Enterprise Agent Platform. Developers can experiment with it in a live demo environment provided by Browserbase.

Early users are already utilizing this technology for automation-intensive tasks such as continuous software testing and managing enterprise knowledge processes. Google has also made reference implementations available on GitHub to facilitate faster development and integration.

The integration of computer use in mainstream AI models marks a crucial point for both innovation and security. While it offers significant automation benefits, it also necessitates caution regarding unauthorized actions, data vulnerabilities, and AI-driven exploitation.

Organizations adopting AI-based agents must prioritize secure deployment practices and continuously monitor agent activities to address emerging threats effectively.