Google Built a Giant ‘Mirror Dimension’ Internet to Test Its Chrome AI Agents

Wired
Read full story· 4 min
Google's Chrome team created a large-scale, controlled alternate internet called the 'Mirror Dimension' to safely train and test AI browsing agents using Gemini models. This simulated web environment mimics real websites and platforms, allowing agents to navigate and resist attacks without risks of uncontrolled behavior on the open internet.

What happened

Google's Chrome team has built a large, isolated web environment called the 'Mirror Dimension' to train and test AI browsing agents powered by Gemini models. This simulated internet includes a wide range of websites and platforms that closely resemble real online services. The environment allows the agents to perform tasks and face security challenges without any risk to the actual internet. Google revealed this project in a blog post and plans to make it available to more teams within the company.

Key details

  • The 'Mirror Dimension' is a controlled, private web ecosystem created by Google's Chrome team.
  • It includes simulated versions of social media, shopping, and banking sites.
  • AI agents are tested for their ability to navigate sites and resist prompt-injection attacks.
  • The environment was built using automated agents and a custom AI harness to generate realistic web content.
  • Researchers had to modify Chrome and use forged web certificates to enable full agent interaction within the environment.

Why it matters

This development allows Google to safely train and evaluate AI browsing agents without risking unintended actions or security issues on the real internet. Teams working on AI at Google now have access to a more realistic and secure testing ground, which can improve the reliability and safety of future AI-powered browsing features. This change reduces the risk of uncontrolled AI behavior affecting real users or web services.


More in Agents