New Delhi: Google has created a giant simulated version of the internet, called the Mirror Dimension, to test the security and reliability of its artificial intelligence (AI) browsing agents without exposing real websites, accounts or user data to potentially harmful actions.

Detailed by Google’s Chrome security team in a November 2026 blog post, the environment contains thousands of simulated websites designed to resemble the services people use every day. These include online shopping platforms, social media feeds, banking dashboards, recipe websites and local marketplaces.

The project is intended to help Google’s Gemini-powered Chrome Auto Browse agents navigate websites, complete multi-step tasks and withstand cyberattacks hidden within online content.

The company says testing AI agents directly on the public internet can be risky because an inexperienced system could accidentally submit forms, alter accounts or perform other actions with real-world consequences. The Mirror Dimension provides a controlled environment where researchers can reproduce such situations without affecting actual users.

Why Google needed a fake internet

AI browsing agents are designed to carry out tasks that would ordinarily require people to navigate websites manually. They can open pages, follow links, scroll through information, enter details into forms and interact with online services.

However, developing these capabilities requires extensive testing across hundreds of websites and thousands of possible actions.

Using the live internet creates two major challenges. First, websites are constantly updated, which makes it difficult for researchers to repeat tests under identical conditions. Second, an AI agent making an incorrect decision on a real website could cause unintended consequences.

Google also identified a problem known as evaluation awareness, or eval awareness. This happens when an AI model recognises that it is being tested and changes its behaviour because it suspects it is operating in a simulated environment.

For example, an agent may behave more cautiously if it notices an unfamiliar test website or an obviously artificial web address. Such behaviour can make test results less representative of how the system would act during ordinary browsing.

Google therefore wanted a realistic environment in which agents would believe they were navigating familiar websites while all their actions remained inside a controlled system.

How the Mirror Dimension works

The Mirror Dimension operates inside an isolated digital sandbox using gVisor, a technology that provides an additional layer of isolation for applications.

When an AI agent attempts to visit a familiar website, the request is intercepted and redirected to a locally hosted synthetic version rather than being sent to the public internet.

Importantly, the browser continues to display the expected website address. An agent attempting to visit YouTube, for instance, can navigate to the familiar YouTube address while receiving content from a simulated website inside Google’s testing environment.

This approach helps reduce the likelihood that the AI will recognise the exercise as a test.

Making the arrangement work with Chrome required additional security engineering. The browser has built-in protections designed to identify suspicious connections and prevent interception of encrypted website traffic.

Google’s engineers therefore configured a local proxy to generate digital TLS certificates that are trusted by a custom root certificate authority installed inside the sandboxed browser.

The company emphasised that this arrangement depends on Google’s control over the entire testing environment. It is not a technique that ordinary attackers can use against regular Chrome users on the public internet without the necessary privileged access.

Each test also begins from a clean state, ensuring that actions performed during one evaluation do not affect subsequent tests.

AI agents are building the simulated websites

Creating thousands of realistic websites manually would require considerable time and resources. Google has addressed this challenge by using a separate system of AI agents to develop and maintain the simulated web environment.

The process divides website creation into three main roles.

Research agents first examine the task requirements and prepare specifications for each website. For an online shopping platform, this could include product catalogues, search filters, shopping carts, discount codes and multi-step checkout processes.

Builder agents then turn those specifications into functioning websites by writing the necessary code and creating page layouts and navigation systems.

Critic agents inspect the generated websites to identify problems. They use Chrome developer tools to interact with pages, test buttons and forms, examine screenshots, check accessibility features and identify errors that need correction.

The agents can then revise the websites and repeat the testing process until the simulated services behave more like real online platforms.

This method allows Google to create interactive websites with working features rather than relying on static pages that an AI agent could quickly identify as artificial.

Testing hidden prompt injection attacks

One of the most important purposes of the Mirror Dimension is to test how Chrome’s AI agents respond to prompt injection attacks.

Prompt injection is a security threat in which malicious instructions are placed inside content that an AI system reads. The instructions may attempt to override the user’s request and persuade the agent to perform an unauthorised action.

For example, a webpage could contain hidden text telling an AI assistant to ignore its original task and disclose private information or interact with an unrelated service.

Such instructions may not always be visible to a human visitor. They can be embedded in images, webpage code, accessibility labels and other elements that browser agents can interpret.

Google’s testing environment allows engineers to introduce these malicious instructions into simulated websites and observe how the AI responds.

Initially, researchers used unusual URL parameters to insert attack instructions. However, they found that advanced AI models could notice these suspicious additions and recognise that they were being evaluated.

Google subsequently moved the attack identifiers into hidden HTTP request headers. This allowed researchers to modify the content delivered by a simulated website without changing the visible web address.

The approach makes it possible to test different types of attacks across multiple stages of a browsing task while keeping the simulated websites looking normal to the AI agent.

Google says the system can generate over 1,000 attack tests daily

The automated red-teaming process has also increased the scale of security testing.

According to Google, a security engineer working manually could create and test a few dozen realistic attack scenarios in a day. The automated pipeline can generate more than 1,000 test candidates daily for each security engineer.

The company said this represents more than 30 times the output achievable through manual testing.

The increased capacity allows security specialists to spend more time identifying new attack techniques while automated systems handle the repetitive work of generating and refining test scenarios.

However, Google has not disclosed a pass rate showing how often its AI agents successfully resisted these attacks or how frequently they were manipulated. The number of test cases generated should therefore not be interpreted as proof that the agents are already immune to prompt injection.

Is Google’s Mirror Dimension available to users?

The Mirror Dimension is an internal testing environment described by Google’s Chrome security team. The company has not announced it as a consumer product or a publicly available service.

The immediate focus is on strengthening Chrome’s Auto Browse capabilities and evaluating how AI agents behave when they encounter potentially hostile online content.

Google engineers have indicated that the environment could be made available to more research teams within the company. They also expect to continue expanding its realism as AI agents become more capable.

The project highlights a broader challenge for the technology industry. As AI assistants move beyond answering questions and begin interacting with websites, accounts and online services, their ability to recognise malicious instructions and avoid unauthorised actions becomes increasingly important.

Google’s Mirror Dimension offers researchers a way to test those safeguards in realistic conditions without putting real users or websites at risk. Its continued development could also provide a model for evaluating other AI systems that interact with the internet.