Google Set Up a Huge “Fake Internet” to Test AI Agents

· AI · Cem Koyluoglu

The Google Chrome team, while developing AI-based browser features using the company's Gemini models, added extra

TL;DR: Google’s Chrome team built a fully controlled, isolated web ecosystem called Mirror Dimension to safely train Gemini‑powered AI browser agents and prevent uncontrolled behavior. The environment mimics real‑world sites, allowing agents to learn tasks while being tested against web threats. The team plans to share this approach with other Google teams for further research.

Key Highlights: • Mirror Dimension is a massive, isolated web ecosystem that mirrors real‑world sites for safe AI training. • The Chrome team used custom AI‑driven development agents to rapidly build realistic sites and data flows. • Researchers had to “hack” Chrome’s security checks within the controlled environment to train agents. • The trained agents showed no harmful cyber capabilities, and the environment will be shared for future research.


Google’s Gemini‑Powered Browser Agents Require a Controlled Environment

Google’s Chrome team discovered that training artificial‑intelligence agents on Gemini models for web‑based tasks can quickly spiral out of control. To prevent this, they built a massive, isolated “alternative universe” internet that the team fully controls.

Mirror Dimension

The team introduced a new project called “Mirror Dimension,” first detailed in a blog post today. The environment is a broad, varied, but isolated web ecosystem that the Chrome team manages entirely. It contains dozens of sites and platforms that mirror real‑world social‑media sites, e‑commerce storefronts, and banking dashboards, all built with the same components and features as their live counterparts. Browser‑assistant agents must interact with these sites through Chrome just as they would on the open web, completing tasks while being subjected to resilience tests such as hidden request‑injection attacks.

Charles Qian, a software engineer on Chrome’s AI security team, explained: “Because the agents are not pre‑programmed, you cannot anticipate everything they might encounter; the discovery space is vast and full of web threats. Any user can enter any request on any website. Therefore, we must explore this space as comprehensively as possible. We have seen the agent begin to develop a mature reasoning ability to determine what information is truly necessary for a task and when to return to the user for confirmation.”

Building the Environment

Creating and populating Mirror Dimension was a significant challenge. The team needed a system that could be built quickly yet be large and diverse enough to represent the scale of the real web. To achieve this, they assembled armies of web‑site and web‑application development agents guided by a custom AI framework. This approach enabled rapid production of realistic web services and data flows, down to details such as entering a promo code on a shopping site or populating a social‑media page with content.

Even though the entire environment is under Google’s control, Chrome remains a browser built with extensive user‑protection and validation checks. For the agents to receive proper training, the researchers had to “hack” the browser and create fake web certificates to bypass these checks. The team stresses that none of these attacks could be performed against normal Chrome users on the open internet. All of this was possible because the team had unlimited access and full control over the entire environment and network.

Safety and Future Access

The researchers report that the tested browser‑assistant agents are extremely harmless and lack any cyber‑capabilities. Nevertheless, they aim to make Mirror Dimension available to other Google teams, providing a safe and controlled test environment for further research.

Doug Turner, Google’s engineering director, said: “We are sharing this approach so that other teams, both inside and outside the company, can consider following it. This is just the first step; we will continue to invest in this approach for training models. As the models become more advanced, we will need to keep working on making this environment increasingly realistic.”

Source: https://www.wired.com/story/google-built-a-giant-mirror-dimension-internet-to-test-its-chrome-ai-agents/

← All tech news