Google Announces Gemini Spark Agent and Antigravity Platform at I/O, Launch Date Not Disclosed

Google announced Gemini Spark at I/O 2026, a personal AI agent that will integrate natively with Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube, and Google Maps. The company positions it as a competitor to OpenAI's agent products, though no general availability date has been provided.

According to Google's FAQ, Gemini Spark runs on two systems: Gemini 3.5 Flash and Antigravity. Antigravity is described as a desktop app, CLI agent tool written in Go, an open source Python SDK wrapping a closed source Go binary, and a VS Code fork.

Security Architecture

Google claims Gemini Spark will execute each task in a fresh, isolated, ephemeral VM on Google Cloud. According to the company's enterprise blog post, traffic routes through an Agent Gateway that enforces Data Loss Prevention policies, with user credentials fully encrypted and never exposed to the agent.

The security model is notable given the sensitivity of data in Gmail and Drive. Security researcher Simon Willison noted the lack of detailed information about prompt injection protections, calling Gemini Spark a potential candidate for what he terms an agent security "challenger disaster" if the isolation proves inadequate.

Antigravity CLI Replaces Open Source Tool

Google announced that the open source Gemini CLI tool (Apache 2.0 licensed TypeScript) will stop working with AI subscription plans on June 18, 2026. The company is replacing it with the closed source Antigravity CLI. The Antigravity SDK Python wrapper is open source, but it bundles a proprietary Go binary.

What This Means

Google is making a significant architectural shift by introducing Antigravity as a foundational platform for agent products, though the rationale for mentioning it alongside Gemini 3.5 Flash in the Gemini Spark FAQ remains unclear. The move from an open source CLI to a closed source alternative represents a notable change in Google's developer tooling strategy.

The lack of concrete security details about prompt injection protections is concerning given the agent's planned access to sensitive user data across Google's productivity suite. The ephemeral VM approach is a standard isolation technique, but its effectiveness against sophisticated prompt injection attacks remains to be demonstrated in production.