In a report published November 20, 2026, Google described a controlled web environment it calls the “Mirror Dimension.” The Chrome team built it to test AI browsing agents powered by Gemini models without letting unfinished agents loose on the open internet.
The environment contains realistic websites and web apps, including social media pages, shopping sites and bank dashboards. Agents use Chrome to navigate them and complete tasks as they would on the real web. The team can also place hidden prompt-injection attacks in the sites to test how the agents respond.
Google says the setup addresses a difficult testing problem: a small simulation cannot capture the variety of the web, while models may behave differently if they recognize that they are being evaluated. Charles Qian, a software engineer on the Chrome AI security team, said the team has seen agents improve at judging what information a task requires and when to ask a user for confirmation. That is the team’s observation, not a claim that every possible web interaction has been tested.
To populate the Mirror Dimension quickly, the team used website-development agents and a purpose-built AI harness to create services and data flows, including shopping promo-code entry and social media content. Researchers also had to work around some of Chrome’s protections—including by creating forged web certificates within the controlled environment—so agents could use the real browser during tests. The team says those techniques depended on its control of the environment and network and would not work against ordinary Chrome users on the open internet.
The researchers say the browsing agents tested so far are benign and lack cyber capabilities. Google’s longer-term aim is to make the environment available to other teams within the company. Doug Turner, Google’s director of engineering, said the approach will need further development as the agents become more sophisticated.




