Google Builds Simulated 'Mirror Dimension' Internet to Test Chrome AI Agents
The sandbox lets Google and outside researchers train and evaluate browsing agents without touching the live web.
Google's Chrome team has built a large, fully controlled simulated internet, which it calls the "Mirror Dimension," to test the AI browsing agents it is developing with its Gemini models. The sandbox contains realistic social media, shopping and banking sites, and agents must complete tasks in real Chrome while facing hidden prompt-injection attacks. Engineers used AI development agents to build the sites quickly, and in doing so had to bypass some of Chrome's protections, such as forging web certificates, which would not be possible against ordinary users. The team says the browsing assistants tested so far are benign and lack cyber capabilities, and it plans to make the environment available to other Google teams and to outside researchers as a safe place to train and evaluate such agents.
Why it matters
The environment gives Google a fully controlled place to probe how Chrome's AI browsing agents behave when they meet hostile content such as hidden prompt-injection attacks. Because the team had to bypass some of Chrome's protections, such as forging web certificates, to run the agents in the real browser, the test setup differs from what ordinary users face. The team says the agents tested so far are benign and lack cyber capabilities, and it plans to share the environment with other Google teams and outside researchers. For readers following the rollout of AI agents in consumer software, the result is a controlled test bed rather than evidence drawn from the open web.
Matters or hype?
Does this story matter, or is it hype? Vote before you see what everyone else said.