The Chrome team has built an isolated copy of the internet called Mirror Dimension to safely test AI agents that work in a browser. It includes social media sites, online stores and banking dashboards designed to resemble the real thing. The agents perform tasks in Chrome and encounter hidden prompt attacks. Google wants to make the environment available to research teams inside the company and, over time, make it more realistic.
Google builds a copy of the internet to test Chrome AI agents
An internet for training agents
Agents that work with websites need to navigate the open internet and withstand web attacks. Tests in a restricted environment provide only part of the experience they need. A model may also recognize that it is being tested and behave differently than it would during ordinary use. But sending unprepared agents onto the open internet is risky, too.
That is why the Chrome team created Mirror Dimension, a varied but isolated web environment under Google's full control. The company first described it in a blog post. It contains websites and platforms, including social media, online stores and banking dashboards. They reproduce the structure and features of real web services, so agents must complete tasks in Chrome just as they would on ordinary websites. During testing, they also encounter hidden prompt attacks designed to test their defenses.
Charles Qian, a software engineer on Chrome's AI security team, explained that it is impossible to predict everything an agent will encounter, and that the web is hostile. A user can enter any prompt on any website, so a broad range of situations needs to be tested. According to Qian, agents are getting better at judging what information a task actually requires and when they should stop and ask the user for confirmation.
How the environment was built
Creating Mirror Dimension was a project in its own right: the environment had to grow quickly enough to capture the scale and diversity of the real internet. The Chrome team used numerous AI agents to build websites and web applications. A dedicated tool coordinated their work on realistic services and data flows, down to entering promo codes in online stores and populating social media pages.
To let agents work in actual Chrome, the researchers had to bypass some of the browser's mechanisms. This was necessary for accurate training: Chrome has many checks and protections that interfere with running it in an artificially created environment. For example, the team created fake web certificates to pass authenticity checks and give agents an ordinary browser experience.
Google stresses that these workarounds would not work against ordinary Chrome users on the open internet. They were possible because the researchers had privileged access: the company controlled the entire environment and network.
The website-assistant agents being tested are harmless and do not have cyberattack capabilities, the researchers say. Google plans to make Mirror Dimension available to other teams inside the company so more researchers can test systems in a safe, controlled environment.
Doug Turner, a director of engineering at Google, said the company is sharing its approach so that both internal and external teams can consider it. He called the current work a first step and said Google would continue investing in this way of training models. As models become more complex, he said, the environment will also need to become more realistic.
Dataist's take: Mirror Dimension lets Google test browser agents against complex scenarios without sending them onto the open internet. For now, Google says the environment will be available to teams inside the company and that work on it will continue; no broader launch has been announced.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X