Google builds simulated 'Mirror Dimension' internet to train Chrome AI agents
Google's Chrome team has created an isolated, fully controlled replica of the internet—complete with fake social media, shopping and banking sites—to train and test its Gemini-powered browsing AI agents. The environment lets agents practice real-world tasks and face simulated attacks like prompt injection without the risks of operating on the live web.
GoKawiil's interpretation of the reporting above, not reported fact.
Training AI agents purely in the real world risks exposing them to genuine security threats and unpredictable scenarios, while narrow test environments can let models detect they're being tested and behave differently—an effect Google calls 'eval awareness.' Building a large, realistic but sandboxed internet suggests Google sees robustness testing against attacks as a prerequisite before letting autonomous agents loose on the actual web.
- Google built an alternate, fully controlled internet called 'Mirror Dimension' to train Chrome's AI browsing agents.
- The simulated environment includes fake social media, shopping, and banking sites with realistic attributes.
- Agents are tested against resilience challenges like prompt-injection attacks to improve security before real-world deployment.
Source: wired.com — Lily Hay Newman, 2026-11-19
Published there as: “Google Built a Giant ‘Mirror Dimension’ Internet to Test Its Chrome AI Agents”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.