Google Built a Fake Internet to Train Chrome AI Agents
Original: Google Built a Giant ‘Mirror Dimension’ Internet to Test Its Chrome AI Agents
Why This Matters
Shows the operational scale required to safely deploy AI agents on the open web.
Google's Chrome team built a controlled, full-scale fake internet called 'Mirror Dimension' to safely train Gemini-powered browsing agents, complete with social media, banking, and shopping sites—and hidden adversarial attacks.
Training AI agents that browse the web creates a classic dilemma: sandboxed test environments are too limited, but unleashing unfinished agents on the real internet is dangerous. Google's answer is the 'Mirror Dimension'—a vast, isolated web ecosystem revealed for the first time in a November 19 blog post.
The environment replicates the breadth of the open web, including social media platforms, shopping sites, and bank dashboards built with the same components as real services, down to promo code entry and populated social feeds. Chrome AI agents must complete tasks exactly as they would on the live internet, while also facing planted prompt-injection attacks to stress-test their defenses.
'Agents are not programmed so you can't predict every possible thing an agent might see,' said Charles Qian, a Chrome AI security engineer. 'We've seen the agent start to develop a mature sense of judgement when it comes to knowing what information is actually needed for a task and when to punt back to the user.'
Building the Mirror Dimension itself required fleets of web-development agents guided by a purpose-built AI harness to rapidly generate realistic sites. Because real Chrome was needed for accurate training, the team had to forge web certificates and effectively 'hack' their own browser—steps they stress would not be possible against regular users on the live web.