Press "Enter" to skip to content

Google creates a parallel internet to train and test Chrome’s AI agents

Google has built a large-scale simulated web environment, dubbed the “Mirror Dimension”, to train its Chrome AI browsing agents without exposing real websites and users to potential risks.

The isolated ecosystem recreates the complexity of the open internet, allowing AI agents powered by Google’s Gemini models to navigate websites, interact with online services and complete tasks as they would in real-world conditions.

The environment includes replicas of social media platforms, online shopping services and banking dashboards. It also exposes agents to security threats such as hidden prompt-injection attacks, testing whether they can recognise malicious instructions and avoid unsafe actions.

Testing AI agents in a hostile web environment

Google says conventional test environments are insufficient because AI models can recognise when they are being evaluated and behave differently. Allowing insufficiently trained agents to operate directly on the internet, meanwhile, could expose users and websites to security risks.

To address this challenge, the Chrome team used AI-powered development tools to build a wide range of realistic websites and applications, complete with functioning data flows and interactive features.

Researchers also modified the controlled testing environment to allow agents to use the standard Chrome browser. This involved bypassing certain browser security checks, including the use of specially created web certificates. Google stresses that these techniques depended on privileged access to the simulated network and could not be used against ordinary Chrome users.

Google plans wider access to the system

According to the Chrome security team, testing has helped agents develop better judgement about which information is relevant to a task and when they should seek user confirmation rather than proceed independently.

Google says it intends to make the Mirror Dimension available to other teams within the company, allowing researchers to test increasingly sophisticated AI systems in a controlled environment.

The company also hopes the approach will encourage other developers to build similar testing systems as AI agents take on more complex tasks online.

Be First to Comment

Leave a Reply