All news
AI, LLM & Agents·18 days ago

Gemini 3.5 Flash adds built-in computer use for agents

Google (The Keyword)Summarized by the ORA·tech AI assistant
Gemini 3.5 Flash adds built-in computer use for agents

Google made computer use a built-in tool inside Gemini 3.5 Flash: developers can build agents that see, reason and act across browser, mobile and desktop via the Gemini API and Gemini Enterprise Agent Platform. Previously a standalone 2.5 model, it is now native to the main Flash model, with two optional enterprise safeguards against prompt injection.

Google announced that computer use is now a built-in tool in Gemini 3.5 Flash — the main Flash model, no longer a separate 2.5 model as before. With it, developers can build custom agents that see the screen, reason and take action across browser, mobile and desktop environments. Google calls it its best performance yet for agentic computer use tasks.

Key points

  • Native in Flash: computer use was previously a standalone Gemini 2.5 model; it is now integrated into Gemini 3.5 Flash — the same model already strong at function calling and Search/Maps grounding.
  • Cross-platform: agents can see, reason and act across browser, mobile and desktop.
  • Target tasks: long-horizon and enterprise automation such as continuous software testing and knowledge work across professional applications.
  • How to use: via the Gemini API and Gemini Enterprise Agent Platform; a Browserbase-hosted demo and an open reference implementation on GitHub are available.
  • Safety: Google uses targeted adversarial training to reduce prompt-injection risk, plus two optional enterprise safeguards: require explicit user confirmation for sensitive/irreversible actions, and automatically stop tasks when an indirect prompt injection is detected.
  • Guidance: a "defense-in-depth" approach — combine secure sandboxing, human-in-the-loop verification and strict access controls.

FAQ

How is this different from the old computer use? It was a separate Gemini 2.5 computer use model; now the capability ships inside the main 3.5 Flash model, avoiding multi-model stitching.

How do developers start? Call it via the Gemini API or Gemini Enterprise Agent Platform; Google provides docs, a GitHub reference implementation and a Browserbase-hosted demo.

Bottom line

Building computer use into the Flash model lowers the bar for agents that operate real UIs — but brings prompt-injection risk that must be managed with user confirmation, sandboxing and access controls.

Summarized from Google's announcement. See the original for full details and availability.

Source
Google (The Keyword)
Read the original
#Gemini#AI agents#computer use#Gemini API#prompt injection
This summary was written by the ORA·tech AI assistant. Read the original for full context.

Related

Chào bạn, mình là Nova. Có phần nào bạn muốn hiểu kỹ hơn không? Cứ nói với mình nhé.