Google describes Mirror Dimension as a synthetic web environment for evaluating Chrome browser-use agents, including Auto Browse. Headless Chrome runs inside an isolated gVisor sandbox, where a local proxy routes requests for familiar URLs to generated sites rather than the public internet. The design pairs contained test runs with realistic pages intended to make the evaluation resemble ordinary browsing. Google’s engineering account describes how the system works.

Useful agent testing has to keep real-world side effects out of the run while making the test feel enough like ordinary browsing to be informative.

What Mirror Dimension is designed to test

Mirror Dimension is designed to evaluate browser-use agents on multi-step web tasks. Google says each test trajectory starts from a clean, identical baseline, so side effects from one run do not carry into the next.

The design addresses two different concerns. Isolation and resets limit exposure to real websites and variation between runs. Familiar URLs and realistic, stateful pages are intended to reduce evaluation awareness: an agent changing its behavior because it recognizes that it is being tested. Neither design goal, by itself, establishes how an agent will perform across the open web.

How the proxy keeps familiar URLs inside the sandbox

A Chrome browser window shows a red “Not Secure” indicator and a “Your connection is not private” warning for YouTube.

In Google’s described setup, headless Chrome runs in a gVisor sandbox. A local proxy intercepts browser requests and serves synthetic pages, keeping them within the test environment. The browser can therefore navigate familiar web addresses without those requests reaching the public sites they resemble.

The proxy also handles HTTPS certificate validation. It creates TLS leaf certificates linked to a custom root certificate authority installed in the sandbox, allowing the controlled browser to validate the intercepted connections.

How Google’s synthetic test sites are built

Google describes thousands of simulated sites based on familiar journeys, such as service marketplaces, checkout flows, social feeds and recipe pages. A person defines the high-level journey; research agents specify the site, builder agents create it with HTML, CSS and JavaScript, and critic agents review the result.

Those reviews cover accessibility semantics, form validation, JavaScript console errors, element boundaries and responsive layouts. Chrome DevTools for Agents lets the agents inspect rendered pages and interact with them by clicking, typing and navigating. The result is a set of interactive sites that can support multi-step tasks, rather than static pages alone.

How the tests exercise prompt-injection defenses

Google says its red-team harness can place test payloads in images, text, accessibility labels and WebMCP responses. These are different ways an agent might encounter indirect prompt injection: instructions embedded in content or tool responses it processes, rather than entered directly as its task.

Payload identifiers travel in custom HTTP headers, which the sandbox filters before serving modified pages. That keeps the test identifiers out of the visible URL while the agent encounters the altered content.

Google reports generating more than 1,000 candidate tests per security engineer per day, over 30 times the few dozen realistic attacks it says an engineer can test manually. That figure describes candidate-generation throughput, not a measured rate of successful defenses. Google’s described focus is browser-use agents working with web primitives and tool calls; it presents evaluation of general-purpose coding agents as a possible extension.