Vietnamese crab exporter

Google built a giant fake internet called Mirror Dimension to safely test its AI agents

Google has built a giant fake internet called the Mirror Dimension to safely train Chrome's AI browsing agents and defend them against hidden cyberattacks without putting real user data at risk.

Advertisement
Google’s Mirror Dimension safely tests Chrome AI agents on a fake internet.

Imagine an entire internet filled with shopping sites, social media feeds, and bank dashboards, except none of it is real. That is exactly what Google has built to safely train the new AI browsing agents coming to Google Chrome.

Detailed on Google’s security blog, this alternate-universe web is called the Mirror Dimension. It works as a massive, locked-off playground where Chrome’s Gemini-powered "auto browse" feature can click links, fill out forms, download podcasts, and shop online without ever touching a real website or putting actual user data at risk.

advertisement

Why Google could not use the real internet

Training an AI agent to handle everyday web errands requires running thousands of tests across hundreds of websites. Doing this on the live internet creates big problems. Real websites change constantly, which makes it hard for engineers to repeat tests accurately.

Worse, letting an unrefined AI loose on the open web is risky, because one mistake could mess up a real account or cause actual harm.

There is also a tricky issue called "eval awareness." Modern AI models are so perceptive that if they figure out they are inside a test, they start acting unusually careful, ruining the results. To see how the AI truly behaves, Google had to trick it into thinking it was browsing the real world.

How the fake internet fools Chrome AI

To pull off this illusion, Google built the Mirror Dimension inside a tightly sealed digital sandbox using a tool called gVisor. Whenever the AI tries to visit a popular website, a local system quietly sends it to a fake copy instead. The clever part is that the web address in Chrome’s bar stays completely normal.

Because Chrome has strict built-in security checks, engineers had to use custom, forged web certificates inside the sandbox so the browser would not flash warning signs.

Google stresses that hackers could not do this to regular Chrome users on the open internet, as it only works because the company has total control over its private test network. Every test also starts from a clean slate so previous actions never affect the next run.

AI agents are building the fake web

Building thousands of realistic websites by hand would be impossible, so Google put teams of other AI agents to work. Research agents first plan what a site needs, such as working shopping carts and promo code boxes.

Builder agents then write the actual website code, while critic agents use Chrome developer tools to click buttons, take screenshots, and send glitches back for fixes. Together, they have built an entire fake web of online stores, recipe sites, local marketplaces, and banking dashboards.

Fighting hidden attacks and what comes next

A major goal of this project is protecting users from "prompt injection" attacks, where hackers hide secret, harmful instructions inside webpage text or images to hijack an AI.

Early on, Google’s AI spotted these tests by noticing strange web addresses, so engineers began hiding attack triggers inside invisible web headers instead. This automated setup lets one security engineer create over 1,000 attack tests daily, more than 30 times what a human could make manually though Google did not share how often its AI passed or failed.

advertisement

Chrome AI security engineer Charles Qian noted that the open web is hostile, but the agent is already developing mature judgement on when to pause and ask users for confirmation.

While the tested agents are harmless and lack hacking skills, Google engineering director Doug Turner said that the company will keep upgrading this fake web to make it even more realistic as AI gets smarter.

- Ends