Artificial intelligence

Google Builds a Fake Internet to Train AI Agents


Credit: AI

Google has built an entire artificial version of the internet to train and test its AI-powered browsing agents. Called Mirror Dimension, the isolated environment recreates thousands of realistic websites, including online banks, shopping platforms and social networks, allowing AI systems to interact with the web without putting real users or businesses at risk.

The project was detailed by Google’s Chrome AI Security team on November 19, 2026. Its main purpose is to test how AI agents navigate websites, complete complex tasks and respond to malicious instructions hidden in online content.

A fake internet that behaves like the real one

Mirror Dimension is much more than a collection of static demo pages. The simulated websites contain interactive features, forms, shopping carts, user accounts and realistic data flows.

For example, an AI agent can browse a fictional banking dashboard, navigate a social media feed or complete an online shopping process. Everything happens inside a controlled environment, with no access to real accounts or financial information.

Google uses a specially designed network system to redirect browser requests to these synthetic websites. This allows agents to use Chrome in a way that closely resembles ordinary internet browsing, even though they never reach the actual websites.

Protecting AI agents from online manipulation

One of the biggest reasons behind the project is a growing security problem known as prompt injection.

This technique involves hiding malicious instructions inside website content, images or other page elements. An AI agent may interpret these instructions as legitimate commands and perform actions that contradict the user’s original request.

With Mirror Dimension, Google’s engineers can deliberately introduce these attacks into simulated websites and observe how AI agents respond.

The company has also developed an automated system that uses specialized AI agents to create and review websites. A research agent defines the requirements, a builder agent develops the site, and critic agents test its functionality and identify problems.

According to Google, this approach can generate more than 1,000 potential attack-test scenarios per security engineer each day, significantly increasing the scale of security testing.

Why Google’s parallel internet matters

As AI assistants evolve into autonomous agents capable of shopping, managing online accounts and navigating websites on behalf of users, security becomes increasingly important.

Allowing unfinished AI systems to experiment on the public internet could expose real users to unwanted purchases, data leaks and other unintended consequences.

Mirror Dimension provides Google with a realistic testing ground where mistakes can happen without affecting the outside world.

The company sees the project as an evolving platform that will become more sophisticated as AI agents gain new capabilities. Google also hopes its approach will encourage other research teams to develop similar environments.

Ultimately, the goal is simple: make AI agents capable of using the real internet while ensuring they can distinguish legitimate user instructions from malicious content.

Leave a Reply

Your email address will not be published. Required fields are marked *