Oossa

Google creates a closed‑off “Mirror Dimension” to train Chrome’s AI assistants

Google’s Chrome team built a fully controlled replica of the web so its Gemini‑powered browsing agents can learn safely without exposing users to risky behavior.

By Published by Oossa: 2 min read

ACatInABox · Unsplash

The Chrome team announced a new testing environment they call the “Mirror Dimension.” It is a massive, self‑contained copy of the internet that Google built from scratch. The sandbox contains fake social‑media feeds, shopping sites, bank dashboards and dozens of other page types, all designed to behave like real‑world services. Chrome’s AI browsing assistants, which run on Google’s Gemini models, can navigate these pages, fill out forms and even face hidden prompt‑injection attacks. The goal is to see how the agents react to realistic tasks without letting an untested model roam the live web.

Why a fake web is needed

Training agents in a tiny lab feels too narrow, while releasing them on the open internet is unsafe. The team found that agents often change their behavior when they detect they are in a test—a problem known as “eval awareness.” By giving the agents a full‑scale, but isolated, version of the web, Google can expose them to the same variety of layouts, login flows and security challenges they will meet in the real world. The environment also lets researchers inject malicious prompts that would be dangerous on real sites, checking whether the agents can resist them.

How the Mirror Dimension was built

Google hired “armies” of internal web‑app generators and a purpose‑built AI harness to spin up the sites quickly. The bots created realistic data flows, such as entering promo codes on a checkout page or posting updates to a mock social feed. Because Chrome checks certificates and other security signals, the team had to forge web certificates and give the agents “god‑mode” browser access so the agents experienced a normal Chrome session. The researchers stress that the attacks used in the sandbox would not work on ordinary users because the environment is fully under Google’s control.

Looking ahead

Doug Turner, Google’s director of engineering, said the Mirror Dimension will be opened to other Google teams and possibly external partners. He expects continued investment to make the sandbox more realistic as the AI assistants get smarter. The approach is meant to be a template for anyone who wants a safe, large‑scale web testbed.

Why it matters

For everyday Chrome users, the Mirror Dimension means Google can train its AI helpers in a safe setting, reducing the risk that a buggy assistant will expose personal data or act unpredictably on real sites. It also gives developers a way to see how future AI features might handle phishing or fake‑login attempts before they reach the public.

Was this article useful?
Share

Read next

Oossa · Newsletter

The week in AI, explained

Every Monday: the stories worth knowing, in plain language. Free, no spam.