Google builds isolated 'Mirror Dimension' web to train and test Chrome AI agents
Google's Chrome team built the Mirror Dimension, an isolated web ecosystem it fully controls, to train and test Gemini-based browsing agents without exposing them to the open internet. Google is detailing the project in a blog post, according to Wired.
Google's Chrome team has built a fully controlled, isolated copy of the web, called the Mirror Dimension, to train and test AI browsing agents built on its Gemini models. Wired reports that Google is detailing the project publicly for the first time in a blog post.
What the Mirror Dimension is
The Mirror Dimension is a sweeping and diverse web ecosystem that the Chrome team entirely controls. It includes social media sites, shopping sites and bank dashboards, all built with the components and attributes of real websites. Browsing assistant agents must interact with these sites and complete tasks in Chrome the way they would on the real internet.
The agents also face resilience tests along the way, including hidden prompt-injection attacks, to probe their defenses.
Why Google built it
According to Wired, the team's reasoning has three parts:
- "Web-use" agents need to be familiar with the open internet, navigate real websites and platforms, and be hardened against web attacks. Training in limited test environments is only so effective.
- Models can develop "eval awareness," behaving differently because they have determined they are in a test or simulation.
- Releasing unrefined agents onto the open internet is a bad idea.
"Agents are not programmed so you can't predict every possible thing an agent might see, the exploration space that we have to address is huge and the web is hostile," said Charles Qian, a software engineer on the Chrome AI security team. He added: "We've seen the agent start to develop a mature sense of judgement when it comes to knowing what information is actually needed for a task and when to punt back to the user for a confirmation."
How it was built
The environment needed to be vast and varied enough to represent the scale of the real web while being developed quickly. The Chrome team used armies of website and web app development agents, guided by a purpose-built AI harness, to create realistic web services and data flows. Details include entering promo codes on a shopping site and populating social media pages.
The agents run in real Chrome, which the team describes as crucial for accurate training. Because Chrome has extensive user protections and validity checks, the researchers had to essentially hack the browser and take steps such as forging web certificates so the agents could have a normal experience.
The team stresses that none of these attacks would be possible against regular Chrome users on the open internet. They were enabled by the god-mode access the team has because it controls the entire environment and network.
Scope and next steps
The researchers say the browsing assistant agents they are 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 can use the safe, controlled environment.
"We're sharing this approach so other teams both internally and externally think about doing it this way," said 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. And I would imagine as they become more sophisticated we'll have to continue working on this environment to make it more and more realistic."
No model names, benchmark scores or pricing accompany the announcement. It describes training and evaluation infrastructure, not a model release.
What this means
The Mirror Dimension addresses two problems the Chrome team names directly: agents trained in narrow test environments, and agents that behave differently when they detect a test. Because Google controls the whole environment, it can run real Chrome against realistic sites and inject prompt-injection attacks without risk to real users or services. Qian says the agent has begun to show judgement about what information a task needs and when to hand back to the user for confirmation. Turner says Google expects to keep making the environment more realistic as models become more sophisticated.
Related Articles
Google rolls out computer use in the Gemini app for Mac for multi-step desktop tasks
Google announced on November 19, 2026 that computer use is rolling out in the Gemini app for Mac. Gemini can carry out multi-step work across local files, desktop applications and websites, working in an isolated background window with a live preview in the chat.
Google previews Tool Retrieval in Gemini Interactions API to load tool schemas only when needed
Google announced Tool Retrieval on November 19, 2026, available in preview in the Gemini Interactions API. Agents can keep large tool libraries on standby and load full parameter schemas only when the current task needs them.
Google adds "temporary threads" to AI Mode in Search, detaching chats from your account
Google is adding "temporary threads" to AI Mode in Search, a chat option that detaches the session from the user's account. The conversations won't resurface in search history or recommendations. The feature is rolling out globally on desktop and mobile across all supported languages.
Gemini Notebook now builds interactive websites from your notes with a "make me a site" prompt
Google's Gemini Notebook, formerly known as NotebookLM, can now turn a notebook of research into an interactive website from a text description. TechRadar tested it on King Arthur research and got a finished site in about five minutes, though the reviewer says claims need checking against sources.
Comments
Loading...