Google built a closed web environment called Mirror Dimension to train Gemini-powered browsing agents. This digital replica of the internet lets the company test how agents handle tasks without using the open web. Software engineer Charles Qian said, "the agents have improved at determining what information a task requires."
Google's Chrome team has developed a closed web environment, dubbed Mirror Dimension, to train Gemini-powered browsing agents. The new ecosystem allows the company to test how these agents handle real-world websites and tasks without exposing them to the open web. The development comes after small test setups were found insufficient for this training due to eval awareness, when models realize they are being evaluated and alter their behavior accordingly.
The Mirror Dimension is a digital replica of the real internet, complete with social media platforms, shopping sites, and banking dashboards.
Each platform has been designed to look and function like its real-world counterpart.
The agents work within Chrome and perform tasks as they would online.
Along the way, they are subjected to security tests such as hidden prompt-injection attacks.
The Chrome AI security team's software engineer Charles Qian told WIRED that the agents have improved at determining what information a task requires and when to seek user confirmation.
To build at scale, the team employed groups of website and app development agents guided by a custom AI harness.
These agents created working services and data flows, from promo code fields on shopping pages to populated social media profiles.
To run standard Chrome inside the setup, researchers had to bypass the browser's built-in security checks, including by creating fake web certificates.
Google said these techniques depend on complete control of the network and can't be used against regular Chrome users.
The company has also stressed that the agents tested so far don't have any cyber capabilities.
Google plans to open Mirror Dimension to other teams within the company.
Doug Turner, Google's director of engineering, said they will continue to invest in this approach and make the environment more realistic as models improve.
The major goal is to keep users safe from prompt injection attacks where hackers embed secret, harmful instructions in webpage text or images to hijack an AI.
