KEY POINTS
- Google's Chrome team built an isolated web ecosystem called the Mirror Dimension to test Gemini-based browsing agents.
- The environment includes realistic social media, shopping and banking sites, and tests agents with hidden prompt-injection attacks.
- Google plans to open the environment to more teams and keep improving its realism as models advance.
Google has built a large, self-contained copy of the web to test the AI browsing agents it is developing for Chrome. According to WIRED, which reported on the project on 19 November 2026, the internal environment is called the Mirror Dimension, and Google is describing it publicly for the first time in a blog post.
The Chrome team is building browsing features on top of its Gemini models. Agents that act on the web need to know how real sites and platforms work and must be hardened against web-based attacks. Limited test setups only go so far, WIRED reported. Releasing immature agents onto the open internet, meanwhile, carries obvious risks.
Another concern is what researchers call eval awareness. This happens when a model detects that it is in a test or simulation and changes its behavior as a result. A realistic, wide-ranging environment is meant to reduce that problem.
The Mirror Dimension contains many kinds of sites, including social media, shopping and bank dashboards, all built with the components of real websites. Agents have to complete tasks in Chrome as they would online. Along the way they face resilience checks such as hidden prompt-injection attacks, which try to slip malicious instructions into content an agent reads.
Charles Qian, a software engineer on the Chrome AI security team, told WIRED that agents are not programmed, so every possible input cannot be predicted. He described the web as hostile and the space to explore as huge. He also said the agent has begun to show a more mature sense of judgment about what information a task needs and when to hand control back to the user for confirmation.
To build the environment quickly, the team used groups of website and web app development agents guided by a purpose-built AI harness. These generated realistic services and data flows, down to details like entering promo codes on a shopping site or filling in social media pages.
Because Chrome has many user protections and validity checks, the researchers had to effectively hack the browser, including by forging web certificates, so agents could use the real Chrome in this closed network. The team stressed that such attacks would not work against ordinary users on the open internet, since they rely on the full control the team has over the environment.
The researchers said the agents under test are benign and have no cyber capabilities. Their aim is to make the Mirror Dimension available across Google. Doug Turner, Google's director of engineering, said the company is sharing the approach so other teams inside and outside Google consider it, and that it is a first step with continued investment expected as models grow more sophisticated.
