Google’s Chrome team is developing AI browsing features powered by its Gemini model and has created an isolated, controlled version of the internet to train and test browsing agents more safely. Called the Mirror Dimension, the environment reproduces websites and online services—including social media, shopping and banking platforms—while allowing researchers to study how agents navigate the web and defend against attacks such as hidden prompt injections.
A controlled alternative to the open internet
A web-usage agent must understand how to navigate real websites and platforms while also defending itself against web-based attacks. Training in a limited test environment can help, but models may develop “evaluation awareness,” changing their behavior when they believe they are being tested or operating in a simulation. At the same time, allowing inexperienced agents to operate freely on the open internet creates unacceptable risks.
Google first detailed the Mirror Dimension in a blog post today. The system is a broad but isolated web ecosystem controlled by the Chrome team. It contains a large range of websites and platforms built with the components and characteristics of real online services, including social networks, shopping sites and banking dashboards.
Browsing Assistant agents must use Chrome to interact with these services and complete tasks as they would on the real internet. Researchers then test the agents’ resilience, including their ability to withstand hidden prompt-injection attacks.
Testing agents against unpredictable behavior
“Agents are not programmed, so we can’t predict everything they might see. The research space we have to deal with is vast, and the web is hostile,” said Charles Qian, a software engineer on the Chrome AI security team. “Any user can ask any prompt on any website, so you need to research that area as thoroughly as possible. And we’re seeing agents start to use more mature judgment in knowing what information they actually need for the task and when to punt back to the user for confirmation.”
Building the Mirror Dimension was itself a significant challenge. Google wanted the environment to be broad and diverse enough to reflect the scale of the real-world web, while also being able to develop it quickly. The Chrome team assembled a fleet of website and web-application development agents, guided by a purpose-built AI harness, to create realistic services and data flows.
The simulated services include details such as entering promotional codes on shopping websites and embedding social media pages. These features are intended to give browsing agents more realistic tasks and interactions during training.
Using real Chrome in a simulated web
Although Google controls the entire environment, Chrome includes extensive user protections and validity checks. To let agents use the real browser—an essential part of accurate training—researchers had to take elaborate measures, including effectively hacking the browser and creating fake web certificates to pass validity checks.
The team emphasized that these attacks would not be possible against ordinary Chrome users on the open internet. They were enabled by the elevated, “Godmode-type” access and privileges available to the teams controlling the environment, network and simulated websites.
Plans to expand access
Researchers said the browsing assistant agents tested so far are harmless and do not have cyber capabilities. Google’s goal is to make Mirror Dimensions available to teams across the company so more researchers can use a secure, controlled environment for testing.
“We’re sharing this approach so that other teams, both internal and external, can consider doing things this way,” said Doug Turner, director of engineering at Google. “This is a first step, and we will continue to invest in this approach to training models. And I think we need to continue to work to make this environment more realistic as the models become more sophisticated.”
Source: www.wired.com


