ThinkFacility Sign in
  1. Home
  2. News
  3. Google

Google tests Chrome's AI agent on a synthetic web, hidden so models can't tell

Agents spotted an odd URL parameter and realized they were being tested, so Google forges certificates to keep the illusion whole.

Google's Chrome security team has published how it tests auto browse, the Chrome feature that lets an AI agent browse the web for you. It keeps the agent off the real web. In a post on Google's security blog, engineers Aaron Selya and Charles Qian describe what they call a mirror dimension: a synthetic web of thousands of simulated sites, run inside a sandbox.

Most of the engineering went into one goal. The agent mustn't be able to tell it's being tested, because the models Google is testing are good at noticing.

Built by
Chrome AI Security
What it tests
auto browse, Chrome's agentic browsing feature
The fake web
thousands of simulated websites inspired by everyday user journeys
Attack tests
more than 1,000 tests candidates every day per security engineer
Against testing by hand
over 30x what one engineer could do

Why the test has to stay hidden

Google calls the problem eval awareness. Hosting every fake site under one test domain gave the game away. So did telling the agent, in the prompt, to go to a sandbox address.

People eating under yellow and blue umbrellas on a sunny patio, in front of a glass office building with a Google sign
Google's headquarters in Mountain View, California. Photo: Jijithecat, CC BY-SA 4.0, via Wikimedia Commons

Now the agent goes to the real address, www.youtube.com for example. A headless Chrome inside an isolated gVisor sandbox catches the connection and hands it to a local proxy, which serves Google's copy instead. That's the person-in-the-middle trick Chrome is built to block, so the proxy mints certificates on the fly from a custom root authority installed only in the sandboxed copy of Chrome.

Even that wasn't enough for the first attack tests. Google's red team passed the ID of each planted attack in a URL parameter, and agents caught on.

We quickly learned that frontier models are far too perceptive for this.

From Mirror Dimension for Agents: Evaluating Browser-Use Agents Without the Web

The fix moved the ID into an HTTP header the agent never sees, so the address bar looks normal.

Agents build the sites for the agent

Nobody hand-codes the fake web. A person describes the user journey Google wants to test, and a research agent turns it into a spec for a site: an online shop with a catalog, a cart and a checkout, say. A builder agent writes it.

Then critic agents go over the result, clicking through the rendered page with Chrome DevTools for Agents and sending problems back. Google says it has a demo video where a fix introduces a fresh bug and the critics catch it, and that the loop puts fidelity ahead of finishing fast.

What it means if you let an AI browse for you

The main threat the sandbox tests for is indirect prompt injection. On the real web an attacker can hide those instructions in an image, a page's accessibility labels or a tool's reply, at any step of a long task. In Google's copy the red team can plant them anywhere it likes.

That matters more each month, because Google's agents are doing more for ordinary users. On November 19, 2026, Google started rolling out computer use in the Gemini app for Mac, which works through apps, files and websites in a background window. It's in English in the US, for Google AI Plus, Pro and Ultra users, and you choose which apps it can touch and approve key actions first.

you choose which apps Gemini can access, watch every step live, pause at any time, and approve key actions before they happen.

From Automate Mac tasks with Gemini computer use

The security post doesn't give success rates for injections against auto browse, so it can't tell you how often one gets through. My own reading is that those approval prompts are worth keeping on for anything that spends money or sends messages.

Selya and Qian write that the same sandbox design could be extended to test general-purpose coding agents with broader toolsets. The post gives no date for that.

More on Google

All Google stories