Google Built a Sealed-Off Copy of the Internet to Test Chrome's AI Agents
Letting browser AI agents loose on the live internet is dangerous, and simple test rigs teach them poorly. The Chrome team built an isolated copy of the web: social networks, stores and banking portals where agents work in a real Chrome. There, they are tested with hidden prompt injections. The takeaway: an agent's test environment should resemble the real web but stay under your control.
AI-processed from Wired; edited by Hamidun News
Imagine teaching an intern to use an online banking portal. Giving them access to real accounts is frightening, and a training mock-up with two buttons teaches them nothing. The Google Chrome team faced roughly this dilemma when it started building browser features on Gemini models. Its solution sounds almost like science fiction: the team built an entire alternative internet that it fully controls. Inside Google it is called the "Mirror Dimension." Wired reported on it in detail, and today Google describes the environment for the first time on its own blog.
Why agents needed their own internet
Agents that browse sites and complete tasks on their own need to know the open web: how real platforms are built, how to navigate them and how to avoid falling for an attack. Restricted test environments give too narrow a picture. There is also a second problem, known as "eval awareness": the model realizes it is being tested and starts behaving differently than it would in real work. And releasing a raw agent onto the actual internet is simply dangerous.
According to Charles Qian, an engineer on Chrome's AI security team, the space of possible situations is huge: agents are not hard-coded, a user can give any request on any site, and the web is hostile. That is why this space has to be explored as thoroughly as possible. Qian says the agent has already begun to develop mature judgment: it understands what information a task really requires and goes back to the user for confirmation at the right moment.
How the "Mirror Dimension" works
It is an isolated but diverse ecosystem of sites: social networks, stores, banking dashboards. They are built with all the components and properties of real sites. The agent has to perform tasks in Chrome just as it would on the real internet. Along the way, it is tested for resilience, for example with hidden prompt injections: instructions concealed on a page that try to make the agent do something the user did not ask for.
The most interesting part is how the environment was populated. It has to be large enough to reflect the scale of the real web, yet it also has to be assembled quickly. So the team enlisted whole armies of AI agents that develop sites and web applications under the control of a purpose-built harness. They produce realistic services and data flows, down to details like entering a promo code in a store or filling out pages on a social network.
Hacking its own browser
For accurate training, agents have to work in a real Chrome, not a stripped-down version. But Chrome is packed with protections and authenticity checks. To give agents a proper experience, the researchers essentially had to hack their own browser and, for example, create forged web certificates for the checks. The team stresses that none of this is possible against ordinary users on the open internet. It works only because the team controls the entire environment and network and has "god mode"-level access.
What comes next and what to take from it
The researchers clarify that the agents under test are harmless and have no cyber capabilities. The goal is to make the "Mirror Dimension" available to other Google teams so that more researchers can work in a safe, controlled environment. Google engineering director Doug Turner says the company is sharing its approach so that other teams, both internal and external, start thinking about this too. This is only a first step, investment will continue, and as models grow more complex, the environment will have to become ever more realistic.
For practitioners, the takeaway is fairly direct. If you are building an agent that clicks through websites, demo pages are not enough. You need an environment that:
- resembles the real web in diversity and small details;
- is fully isolated, so the agent's mistakes harm nothing;
- contains attacks, such as hidden instructions, and checks how well the agent resists them;
- does not give itself away as a test, otherwise you will measure "exam" behavior rather than behavior on the job.
The limitations are obvious too. Google has not yet published any quantitative results: no share of attacks repelled, no comparison with training on ordinary test rigs. All that is known is the engineers' words about the agent's "mature judgment." But the approach itself, in which AI agents build a proving ground for other AI agents, will likely become the standard for anyone who trusts models with banking portals and purchases.
- Nov 20, 2026Google teaches Gemini to run your Mac in the background without taking over the screen or cursor
- Nov 20, 2026Anthropic shares jump 17% on first day after record $81.3 billion IPO
- Nov 20, 2026Google opens Genie 4 world model in its API: one minute of walking costs about $6
- Nov 20, 2026Gemini API Cuts Token Use for a 200-Tool Agent Nearly 7x, by Google's Estimate
Key Questions Answered
Why agents needed their own internet?
Agents that browse sites and complete tasks on their own need to know the open web: how real platforms are built, how to navigate them and how to avoid falling for an attack. Restricted test environments give too narrow a picture. There is also a second problem, known as "eval awareness": the model realizes it is being tested and starts behaving differently than it would in real work. And releasing a raw agent onto the actual internet is simply dangerous.
How the "Mirror Dimension" works?
It is an isolated but diverse ecosystem of sites: social networks, stores, banking dashboards. They are built with all the components and properties of real sites. The agent has to perform tasks in Chrome just as it would on the real internet. Along the way, it is tested for resilience, for example with hidden prompt injections: instructions concealed on a page that try to make the agent do something the user did not ask for. The most interesting part is how the environment was populated.
What are the key details on Hacking its own browser?
For accurate training, agents have to work in a real Chrome, not a stripped-down version. But Chrome is packed with protections and authenticity checks. To give agents a proper experience, the researchers essentially had to hack their own browser and, for example, create forged web certificates for the checks. The team stresses that none of this is possible against ordinary users on the open internet. It works only because the team controls the entire environment and network and has "god mode"-level access.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.