Google has created a synthetic version of the internet, dubbed the "Mirror Dimension," to serve as a secure training ground for its AI browsing agents. This isolated environment allows the AI to practice navigating websites, clicking buttons, and filling out forms in thousands of simulated scenarios without posing a security risk to the real internet. The project aims to ensure the agents are safe, reliable, and resistant to threats before they are deployed to users.
The primary agent being tested is for a Google Chrome feature called "Auto Browse," which runs on the Gemini model and is designed to handle tasks like researching flights or managing subscriptions on a user's behalf. This controlled testing environment is crucial for preventing the AI from developing "eval awareness"—where it knows it's being tested—and avoids accidental interactions with the live web, which has occurred in past AI security tests.