Google’s Chrome team has created the “Mirror Dimension,” an isolated, Google-controlled ecosystem of realistic websites and web applications for training and testing AI browsing agents. The environment is designed to reproduce the complexity and risks of the open internet—including social media, shopping and banking sites, as well as hidden prompt-injection attacks—without exposing unsophisticated agents to real users or services.
A controlled version of the web
A “web usage” agent must understand how to navigate real websites and platforms while defending itself against web-based attacks. Training in a limited test environment can be effective, but models may develop so-called “evaluation awareness,” changing their behavior because 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 presents clear risks. Google says the Mirror Dimension is intended to address both challenges.
Google first detailed the Mirror Dimension in a blog post published today. The system is a broad, diverse and isolated web ecosystem controlled by the Chrome team. It contains websites and platforms built with the components and attributes of real services, including social media sites, shopping systems and banking dashboards.
Browsing Assistant agents interact with those sites and complete tasks in Chrome as they would on the real internet. They are also exposed to resilience tests, including hidden prompt-injection attacks designed to assess how well they defend themselves.
“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 realistic websites and data flows
Creating the Mirror Dimension was itself a significant challenge. The Chrome team wanted the environment to be broad and diverse enough to reflect the scale of the real-world web while also being quick to develop.
To build it, the team assembled a fleet of website and web-application development agents. A purpose-built AI harness guided those agents as they rapidly created realistic web services and data flows, including details such as entering promotional codes on shopping sites and embedding social media pages.
Replicating Chrome safely
Although Google controls the entire environment, Chrome includes extensive user protections and validity checks. To let agents use the real Chrome browser—an essential part of accurate training—researchers had to take elaborate steps to create a normal browsing experience. Those steps included essentially hacking the browser and creating fake web certificates to satisfy validity checks.
The team emphasizes that these attacks would not be possible against ordinary Chrome users on the open internet. Google was able to carry them out because its teams control the entire environment and network, giving them what the source describes as “Godmode-type access and privileges.”
Plans to expand access
Researchers say the browsing assistant agents tested so far are harmless and have no cyber capabilities. Their goal is to make the Mirror Dimension available to teams across Google, giving more researchers access to a secure and controlled testing environment.
“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












