Google Built A Mirror Dimension Internet To Test Its Chrome AI Agents

Google's Chrome security team trains its AI browsing agents inside what it calls the "Mirror Dimension": an isolated, entirely synthetic copy of the internet, detailed publicly for the first time in a Google blog post on Thursday, as reported by Wired.
Why fake the whole web
The problem is specific to web-use agents. They need to be fluent in the real internet, how to navigate actual sites and platforms, and how to resist attacks on the web, so small fixed test environments train them poorly. Worse, models can develop "eval awareness": if a test site is obviously artificial, the agent figures out it is being tested and behaves differently.
But letting unrefined agents act on the live internet is equally untenable. Browser-agent training requires thousands of read and write actions across hundreds of sites, often with the model inferring destinations from a user's phrasing rather than being given URLs, and unrestricted agent actions cause real harm on real services.
The Mirror Dimension splits the difference. It contains a vast array of simulated sites, including social media, shopping, and banking dashboards, built with the components and data flows of real web services. Agents request familiar addresses while a headless Chrome instance runs inside an isolated gVisor sandbox, and a proxy serves the generated site instead. Each test trajectory gets a clean, identical starting point, so side effects from one run cannot leak into the next.
Prompt injections hidden in headers
The environment also works agents through resilience tests, including hidden prompt-injection attacks designed to check the agents' defenses. According to one account of the design, injections are smuggled in places like custom HTTP headers so the agent cannot infer it is under test. Chrome AI security engineer Charles Qian told Wired the exploration space is the core issue: agents are not programmed, any user can prompt anything on any site, so the team has to cover as much of that space as it can.
The construction itself was a project in scale. The team deployed swarms of website-building agents, driven by a custom AI harness, to generate realistic web services and data flows quickly, down to details like promo-code entry on a shopping site.
One wrinkle: Chrome ships with so many security checks that running real agents in real Chrome inside the fake web required the researchers to make forged TLS certificates and similar accommodations, possible only because they control the entire environment and network. Google stresses those tricks would not work against ordinary Chrome users on the open internet.
The eval problem it exposes
Mirror Dimension is the clearest institutional acknowledgment yet that agent evaluation has become a research problem in its own right. Labs cannot test dangerous capabilities on the live web, and toy sandboxes fail because models detect fakery. A generated internet solves both, at the cost of building a second internet.
Google says the browsing agents tested this way are benign and have no cyber capabilities, and plans to open the environment to other teams across the company, and to encourage external adoption. Director of engineering Doug Turner framed it as a first step: as models get more sophisticated, the fake web has to get more realistic.
Sources:
