Google’s Chrome security team built Mirror Dimension, an isolated copy of the web with thousands of simulated sites, so it can test Chrome’s auto browse agent without letting it loose on real websites. The fake sites include online stores with carts, coupon codes and multi-step checkout, and they’re designed by a team of AI agents that write specs, code the pages and check them against real-world web standards. When the agent goes to a familiar address like youtube.com, a proxy inside the sandbox answers with the fake version, so the model doesn’t realize it’s being tested. Engineers use the setup to hide prompt injection attacks in page text, images or code at any point in a task, and Google says automated red-teaming now turns out more than 1,000 test cases per engineer each day.






