Google Creates Mirror Dimension, a Fake Internet to Safely Test Chrome AI Agents

Share Us

38
Google Creates Mirror Dimension, a Fake Internet to Safely Test Chrome AI Agents
21 Nov 2026
min read

News Synopsis

Google has developed Mirror Dimension, a simulated internet that allows Gemini-powered Chrome AI agents to browse websites, complete tasks and undergo security testing without affecting real users.

Google Creates Mirror Dimension, a Fake Internet to Safely Test Chrome AI Agents

Google has developed an artificial version of the internet called Mirror Dimension to train and test its upcoming AI browsing agents in a controlled environment. The system recreates familiar online experiences, including shopping platforms, social media feeds, recipe websites and banking dashboards, without connecting to actual services.

The technology is designed to support Chrome's Gemini-powered Auto Browse feature, which can perform online activities such as clicking links, completing forms, downloading content and shopping.

According to Google's security blog, Mirror Dimension provides a protected digital environment where AI agents can learn to perform web-based tasks without risking real accounts, personal information or financial transactions.

Why Google Created a Fake Internet for AI Agents

Training AI agents to perform everyday internet activities requires extensive testing across thousands of tasks and hundreds of websites.

However, using the actual internet for such experiments creates several challenges. Websites frequently change their layouts, features and functionalities, making it difficult for engineers to reproduce identical testing conditions.

Another concern is the possibility of AI agents making mistakes while interacting with real websites. An incorrectly submitted form, an unintended purchase or an accidental account modification could create problems for users.

The Challenge of AI Test Awareness

Google also identified a problem known as 'eval awareness'. Advanced AI models can sometimes recognise when they are operating inside a testing environment.

When an AI system realises that its actions are being evaluated, it may behave differently from how it would perform during normal internet usage. This can affect the reliability of test results.

Mirror Dimension addresses this challenge by recreating realistic browsing conditions, allowing engineers to observe how AI agents respond to different situations without exposing real websites to potential risks.

How Google's Mirror Dimension Works

Google has built Mirror Dimension inside an isolated digital environment using a security technology called gVisor.

The system creates a simulated version of the internet where AI agents can interact with websites that look and function like their real-world counterparts.

Whenever an AI agent attempts to open a popular website, the internal testing system redirects the request to a locally hosted replica.

Interestingly, the website address displayed in Chrome's address bar remains unchanged, helping maintain the appearance of normal internet browsing.

Special Security Certificates Keep the Simulation Running

Chrome includes strict security checks that normally identify suspicious website connections.

To make the simulated websites work without triggering browser warnings, Google's engineers use specially created web certificates within the private testing environment.

Google has clarified that this setup cannot be replicated against ordinary Chrome users browsing the public internet because the system depends on complete control over its isolated testing network.

Each testing session also begins with a fresh environment. This ensures that actions performed during previous experiments do not influence subsequent tests.

AI Agents Are Helping Google Build the Fake Internet

Creating thousands of realistic websites manually would require enormous amounts of time and resources. To solve this problem, Google is using multiple AI agents to develop and maintain the simulated web environment.

These agents perform different responsibilities, working together to create functional websites and identify technical problems.

Research Agents Plan Website Features

The process begins with research agents that identify the requirements of individual websites. For example, a simulated shopping platform may need product listings, shopping carts, checkout pages and promotional code options.

Similarly, other websites require features such as search functions, interactive menus and account dashboards.

Builder Agents Develop Functional Websites

Once the requirements are prepared, builder agents generate the code needed to create the websites.

These systems develop the actual website structures and interactive features, allowing Chrome's AI agents to perform realistic browsing activities.

Critic Agents Identify Errors and Improve Performance

Critic agents test the websites using Chrome's developer tools.

They interact with buttons, capture screenshots and identify problems that could affect functionality. Their feedback is then used to make corrections and improve the simulated websites.

Through this coordinated approach, Google has created a large collection of virtual online stores, recipe platforms, local marketplaces and banking dashboards.

Mirror Dimension Helps Protect AI Agents From Prompt Injection Attacks

One of the most important objectives behind Mirror Dimension is improving protection against prompt injection attacks.

These attacks involve hiding malicious instructions within webpage content, images or other online elements. Such instructions can attempt to manipulate an AI agent into ignoring its original task or performing unauthorised actions.

Because AI browsing agents interact directly with online content, they could become potential targets for these attacks.

Google Increases the Number of Security Tests

During early experiments, Google's AI agents identified certain security tests because they noticed unusual website addresses.

To make the tests more realistic, engineers started placing attack triggers inside invisible web headers rather than relying on suspicious URLs.

This approach allows security teams to evaluate how AI agents respond to hidden instructions.

According to Google, the automated testing system enables a single security engineer to generate more than 1,000 attack tests daily. This is over 30 times the number that could be created manually.

However, Google has not disclosed the overall success rate of its AI agents in identifying or resisting these attacks.

Google Plans to Make Mirror Dimension More Realistic

Google engineers believe that AI browsing agents need strong judgement when interacting with the open internet.

Chrome AI security engineer Charles Qian highlighted that the internet can contain hostile content and that AI agents must learn when to stop an activity and seek user confirmation.

Google engineering director Doug Turner also indicated that the company would continue improving Mirror Dimension as AI capabilities advance.

The current testing agents are designed for harmless activities and do not have hacking capabilities.

Conclusion

Google's Mirror Dimension represents an important step towards making AI-powered internet browsing safer and more reliable. By combining simulated websites, multiple AI development agents and automated security testing, the company can evaluate potential risks without affecting real users. As Gemini-powered browsing features become more advanced, this virtual internet could play a crucial role in improving AI security, preventing prompt injection attacks and building greater trust in autonomous browsing technology.