Newsletter Subscribe
Enter your email address below and subscribe to our newsletter
[forminator_form id="25163"]

blog.google.blog.google.cybrfunk.com+1.Google has built a sealed-off, synthetic copy of the web called the "mirror dimension" to train and test the AI agents in its Chrome browser. Inside it, browsing assistants can shop, scroll social feeds and use bank dashboards without touching the real internet. Chrome's security team described the system for the first time in a Google blog post published Thursday, November 19, 2026. The same day, Wired reported that Google Alphabet Inc. hopes to share the approach with other research teams.cybrfunk.com+1
The project reflects a problem facing every company building agents. Testing them on the live web is unreliable and risky, but testing them in obviously fake environments can make the results meaningless.
Engineers Aaron Selya and Charles Qian wrote that the environment is used to evaluate Chrome's auto browse feature across thousands of simulated websites modeled on everyday tasks. A central goal is avoiding "eval awareness," where a model notices it is being tested and acts differently. So the agent goes to real addresses like youtube.com. A local proxy inside an isolated gVisor sandbox quietly catches the request and serves a generated site instead, and it creates TLS certificates on the fly so Chrome doesn't flag the connection.blog.google
Google said the team has stressed that this kind of interception is possible only because it controls the whole sandbox. An outside attacker could not do the same to a regular user's device. To fill the fake web, Google used a harness of research, builder and critic agents that design and refine sites from a single prompt. The resulting sites range from online stores with coupon codes to recipe pages.blog.google
The sandbox also serves as a test bed for indirect prompt injection, where hidden instructions in web content try to take over an agent. Google said frontier models were "far too perceptive" for its first attack setup, which put markers in URLs. Attack payloads are now passed through hidden HTTP headers instead. The company says its automated pipeline produces more than 1,000 test candidates per security engineer each day, over 30 times what one engineer could write by hand.blog.google
"The exploration space that we have to address is huge and the web is hostile," Qian told Wired.cybrfunk.com
The news comes a day after Wired reported that Google was reshuffling the team behind Project Mariner, its earlier browser agent prototype. Google said it believes synthetic internets like this one will become standard for agent safety research, and that the design could be extended to coding agents.wired.com+1
"We're sharing this approach so other teams both internally and externally think about doing it this way," Doug Turner, Google's director of engineering, told Wired. "This is the first step."cybrfunk.com