Simon Willison’s blog
Sponsor: Datadog — Deliver reliable AI faster with LLM observability. Read the best practices guide
May 20, 2026
Google I/O is hard to write about this year, because I have a rule: I don’t write about things I can’t try myself, and a lot of the major announcements are just “coming soon.”
I actually prefer writing about features that are already generally available, because I’ve run into cases before where the preview version and the version eventually released to the public turned out not to match.
Aside from Gemini 3.5 Flash, the most interesting announcement was probably Google’s upcoming OpenClaw competitor, Gemini Spark, described as “your personal AI agent” that can “natively connect to your favorite Google apps like Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube, and Google Maps.” Its FAQ also contains this confusing detail:
What model powers Gemini Spark?
Gemini Spark is powered by Gemini 3.5 Flash and Antigravity.
The antigravity.google site currently lists Antigravity as a desktop app, a CLI agent tool written in Go, Antigravity SDK (an open source Python wrapper around a bundled closed-source Go binary), and the original Antigravity IDE, a fork of VS Code.
My guess is that Gemini Spark, the user-facing hosted agent product, probably runs on that Go binary — but I don’t understand why that would be worth putting in the FAQ!
I naturally went looking for information about how Gemini Spark intends to handle prompt injection risk. The best information I could find was from Everything Google Cloud customers need to know after Google I/O, a post aimed at enterprise customers, which says:
Spark runs on Google Cloud in a fully managed, secure runtime environment, which means you get enterprise-grade security without having to manage the underlying infrastructure. Each task runs in a fresh, tightly isolated, ephemeral virtual machine to help ensure that data never overlaps across sessions. To protect your enterprise, all traffic flows through our secure Agent Gateway, which enforces data loss prevention (DLP) policies, while user credentials remain fully encrypted and are never directly exposed to the agent.
Given how many people are going to be passing extremely sensitive data through Gemini Spark, I hope they’ve made this thing truly bulletproof; otherwise it may become the leading candidate for the agent security “challenger disaster” we have yet to see.
Also notable is Google’s announcement in Transitioning Gemini CLI to Antigravity CLI that the open source Gemini CLI tool, a TypeScript project under the Apache 2.0 license, will stop working with its AI subscription plans starting June 18 and be replaced by the new closed-source Antigravity CLI.
Recent articles
- Gemini 3.5 Flash: more expensive, but Google plans to use it everywhere - May 19, 2026
- LLMs over the last six months, in five minutes - May 19, 2026
- Notes on the xAI/Anthropic datacenter deal - May 7, 2026
This is a note from Simon Willison, published on May 20, 2026.
Monthly briefing
Sponsor me for $10 a month to receive a curated email rounding up the month’s most important LLM developments.
Pay me to email you less!