A controlled web for training agents
Google’s Chrome team has built a completely isolated web ecosystem under its own control, dubbed the “Mirror Dimension,” to train AI navigation agents based on Gemini models. Wired reports this, citing a blog post in which the company describes the initiative for the first time. The environment reproduces real sites and platforms—social networks, online stores, and banking dashboards—built with the same components and attributes as authentic services, but with no connection to the open internet.
The choice stems from a technical trade-off indicated by the source. Training agents in limited test environments is ineffective, because models can develop “eval awareness,” meaning they behave differently when they recognize they are in a simulation. Conversely, leaving them free on the open web without adequate training is considered dangerous. The Mirror Dimension aims to offer a context broad and realistic enough to overcome both limitations.
Tasks, attacks, and defenses
In the environment, agents must complete tasks on Chrome as they would on the real web, also facing resilience tests. These include hidden prompt injection attacks, inserted to verify the agent’s defense capabilities. Charles Qian, a software engineer on Chrome’s AI security team, explains that the exploration field is enormous and that the web is hostile: any user can make any request on any site. According to Qian, the tested agents began to develop a more mature sense of judgment in distinguishing the information needed for a task and in asking the user for confirmation when necessary.
Building and populating the Mirror Dimension was a challenge in itself. To quickly obtain a variety comparable to that of the real web, the team used teams of website and application developer agents, guided by a dedicated AI system, to create realistic services and data flows, including details such as entering promotional codes or filling out social pages.
Because Chrome has numerous protections and validity checks, to enable agents to use the real browser—a condition considered crucial for accurate training—researchers had to effectively force the browser, for example by creating forged web certificates. The team stresses that none of these attacks would be possible against normal Chrome users on the open web: they are made possible only by the group’s privileged access, which controls the entire environment and network.
Next steps and open questions
Google says the navigation agents tested so far are benign and do not have offensive cyber capabilities. The stated goal is to make the Mirror Dimension available to internal teams, so that more researchers can use it as a safe and controlled environment. Doug Turner, Google’s director of engineering, describes the approach as a first step, intended to be followed by further investment in model training and an increasingly realistic environment.
Several aspects that the material does not clarify remain to be verified: what quantitative results the tests achieved, how many resources were invested, which internal teams already have access to the platform, and whether the approach will also be adopted outside Google. The source also does not indicate whether the Mirror Dimension will be made available to parties outside the company.
