Google creates 'mirror internet' for testing Chrome AI agents
The Chrome team has developed an isolated web environment that simulates the open internet to train AI agents and test their defenses against attacks.

The 'mirror dimension' includes social networks, stores, and banking dashboards modeled after real websites. Agents perform routine tasks in Chrome and encounter hidden instruction-injection attacks. This allows testing in a realistic environment without releasing unprepared systems into the open internet. In limited simulations, models may recognize they are being tested and alter their behavior.
The environment was rapidly populated by specialized AI tools and agents: they created websites, applications, and data flows, including promo codes in stores and social media posts. To ensure the tested systems operated in the real Chrome browser, researchers had to bypass the browser's protective mechanisms and create fake certificates. Google emphasizes that this was only possible due to full control over the test network and poses no threat to regular users.
The tested agents currently lack cyber capabilities. Google plans to open the 'mirror dimension' to other teams within the company and continue making it more realistic as models evolve.



