Google’s Chrome team built an entire fake internet to test its Chrome AI agents safely. Called the Mirror Dimension, it is an isolated web ecosystem that the team controls completely, and Google is describing it publicly for the first time today.
The Chrome team is developing AI browsing features on Gemini models and needed a place where agents could practice without causing harm.
Why a closed copy of the web
Agents that browse need to know the open internet: how real sites and platforms work, and how to resist attacks. Small test environments only go so far. Models can also develop what researchers call eval awareness, behaving differently once they decide they are in a test. Letting unrefined agents loose on the live web isn’t an option either.
“Agents are not programmed so you can’t predict every possible thing an agent might see,” said Charles Qian, a software engineer on the Chrome AI security team. “The web is hostile,” he added, and any user can give any prompt on any site.
What’s inside it
The Mirror Dimension holds a wide variety of sites built with the components of real ones, including social media, shopping and bank dashboards. Agents must finish tasks in Chrome the way they would on the open web. Along the way they face resilience tests such as hidden prompt-injection attacks.
Building enough sites was a task of its own. The team deployed large numbers of website and web app development agents, guided by custom-built tooling, to create realistic services and data flows quickly. The details go down to entering promo codes on a store or filling social media pages with content.
Using real Chrome in a fake web
Agents needed to run in real Chrome for the training to be accurate. Chrome has many user protections and validity checks, so researchers had to effectively hack their own browser, including forging web certificates, so agents would see a normal experience.
The team stresses that none of this could be done against regular Chrome users on the open internet. It was possible only because they control the whole environment and network, with the kind of god-mode access that regular attackers lack.
Why eval awareness matters
If an agent can tell it’s being tested, its good behavior in the lab says little about how it acts in public. A convincing, varied environment lowers that risk, because the agent faces the same kinds of pages, forms and traps it would meet in real use. Sites that look and behave like real ones also let the team plant attacks that are hard to anticipate, since the pages are theirs to rewrite.
The setup also makes experiments repeatable. When engineers change a model or a defense, they can rerun the same set of tasks and traps and compare the results, something the unpredictable live web can’t offer.
What the approach changes for security
Prompt injection is one of the hardest problems for browsing agents. A page can hide instructions that the agent mistakes for a request from the user, and a real attacker only needs one successful page. By building the pages itself, the Chrome team can write those traps in advance, vary them and see whether the agent falls for them, without exposing anyone to risk.
It also lets the team measure judgment, not just success. A good agent should complete a purchase on a shopping site while declining to share bank details that the task doesn’t need, and it should ask the user before doing something consequential.
What Google has learned
Qian says the agent has started to show a mature sense of judgment about what information a task needs and when to hand control back to the user for confirmation. The researchers add that the browsing assistants under test are benign and have no cyber capabilities.
What happens next
Google wants to open the Mirror Dimension to teams across the company so more researchers can test in a safe, controlled space. Doug Turner, Google’s director of engineering, said the company is sharing the approach so other teams, inside and outside Google, consider doing likewise. “This is the first step,” he said, and he expects continued investment as models become more sophisticated and the environment has to grow more realistic. For anyone building Chrome AI agents, the message is that testing on a controlled copy of the web is becoming a standard step.
Source: WIRED