Google Built a Giant ‘Mirror Dimension’ Internet to Test Its Chrome AI Agents

Google Chrome has built a fully controlled alternate internet to train its AI agents without exposing them to the dangers of the…

By Vane November 19, 2026 2 min read
Google Built a Giant ‘Mirror Dimension’ Internet to Test Its Chrome AI Agents

Google Chrome has built a fully controlled alternate internet to train its AI agents without exposing them to the dangers of the real web.

The team discovered that training browsing assistants on limited test sites was insufficient. Models can develop “eval awareness,” changing their behaviour because they know they are in a simulation. Releasing unrefined agents on the live internet is too risky. The solution was a sprawling, isolated ecosystem that mimics the open web while remaining entirely under Google’s control.

This environment, called the Mirror Dimension, contains a vast array of websites and platforms. It includes social media sites, shopping services, and bank dashboards built with the same components and attributes as real web pages. Agents must interact with these sites and complete tasks within Chrome exactly as they would on the live internet. They also face resilience tests, including hidden prompt-injection attacks, to verify their defences.

“Agents are not programmed so you can’t predict every possible thing an agent might see,” says Charles Qian, a software engineer on the Chrome AI security team. “The exploration space that we have to address is huge and the web is hostile.”

Any user can ask any prompt on any website, so the team needed to explore that space thoroughly. They have observed the agent develop a mature sense of judgement regarding what information is actually needed for a task and when to ask the user for confirmation.

Constructing and populating the Mirror Dimension was difficult. It needed to be vast and varied enough to represent the scale of the real web while being developed quickly. The Chrome team gathered armies of website and web app development agents guided by a purpose-built AI harness to create realistic web services and data flows. This extended to details like entering promo codes on a shopping site or populating social media pages.

Even though Google controls the entire environment, Chrome includes many user protections and validity checks. To allow agents to use the real browser for accurate training, researchers essentially had to modify the browser. They took elaborate steps like creating forged web certificates for validity checks so the agents could have a normal experience.

The team emphasises that none of these attacks would be possible against regular Chrome users on the open internet. These tests were facilitated by the god-mode type access and privilege the team inherently has because they control the entire environment and network.

Researchers say the browsing assistant agents they have been testing are benign and do not have cyber capabilities. Their goal is to make the Mirror Dimension available to teams across Google so more researchers have access to a safe and controlled environment for testing.

“We’re sharing this approach so other teams both internally and externally think about doing it this way,” says Doug Turner, Google’s director of engineering. “This is the first step, there’s going to be continued investment in this approach to train models.”

As the agents become more sophisticated, the team expects to continue working on the environment to make it more realistic.

Scroll to Top