Chrome Mirror Dimension: Google Built a Fake Internet So Agents Cannot Tell
The hardest part of testing a browsing agent turns out to be keeping it from noticing the test. On Thursday at 17:00 UTC the Chrome AI Security team published how it evaluates auto browse, Chrome's agentic browsing feature, inside what it calls the mirror dimension: a hermetic, synthetic copy of the web with thousands of simulated sites. Wired ran the story the same hour. Google's post: https://blog.google/security/mirror-dimension-for-agents/ and Wired: https://www.wired.com/story/google-built-a-giant-mirror-dimension-internet-to-test-its-chrome-ai-agents/
The problem is called eval awareness. If every fake site lives under a sandbox domain, the agent sees a weird URL and starts behaving like it knows it is being watched. So Google lets the agent type the real address, say youtube.com, and inside a gVisor sandbox all traffic is routed to a local proxy that serves a generated site instead. To keep Chrome from throwing certificate errors, the proxy mints certificates that chain to a custom root installed only inside the sandbox. Google stresses this works because it controls the whole machine; an outside attacker could not do it to a normal user without admin access already.
The sites themselves are built by agents. A human writes the user journeys to test. A research agent turns that into a spec, a builder agent writes the HTML, CSS and JavaScript, and critic agents check accessibility semantics, form validation, console errors and layouts, using Chrome DevTools for Agents (the chrome-devtools-mcp server) as eyes and hands. The detail I like most: Google's first version passed test payload IDs through URL parameters, and frontier models were perceptive enough to notice the odd parameter and get suspicious. They moved it out of the visible URL entirely.
The output number is the business case. One security engineer can hand-craft a few dozen realistic prompt-injection tests a day; the pipeline produces more than 1,000 test candidates per engineer per day, over 30x, while humans focus on inventing new attack ideas. Injections can sit in visible text, images, accessibility labels or tool responses such as WebMCP, at any step of a multi-turn task. Charles Qian of Chrome AI Security told Wired the web is hostile and the exploration space is huge.
This is the most concrete answer yet to a question the safety world keeps raising: how do you evaluate a model that can tell it is being evaluated? Google's answer is not a better prompt, it is a world realistic enough that the model has no tell. Doug Turner told Wired Google is sharing the approach so other teams build it this way, and the post says it could extend to general coding agents. Expect hermetic synthetic internets to become table stakes for anyone shipping an agent that touches the web. The catch: the better the mirror, the less anyone outside Google can check what happens inside it.
← Back to all articles
The problem is called eval awareness. If every fake site lives under a sandbox domain, the agent sees a weird URL and starts behaving like it knows it is being watched. So Google lets the agent type the real address, say youtube.com, and inside a gVisor sandbox all traffic is routed to a local proxy that serves a generated site instead. To keep Chrome from throwing certificate errors, the proxy mints certificates that chain to a custom root installed only inside the sandbox. Google stresses this works because it controls the whole machine; an outside attacker could not do it to a normal user without admin access already.
The sites themselves are built by agents. A human writes the user journeys to test. A research agent turns that into a spec, a builder agent writes the HTML, CSS and JavaScript, and critic agents check accessibility semantics, form validation, console errors and layouts, using Chrome DevTools for Agents (the chrome-devtools-mcp server) as eyes and hands. The detail I like most: Google's first version passed test payload IDs through URL parameters, and frontier models were perceptive enough to notice the odd parameter and get suspicious. They moved it out of the visible URL entirely.
The output number is the business case. One security engineer can hand-craft a few dozen realistic prompt-injection tests a day; the pipeline produces more than 1,000 test candidates per engineer per day, over 30x, while humans focus on inventing new attack ideas. Injections can sit in visible text, images, accessibility labels or tool responses such as WebMCP, at any step of a multi-turn task. Charles Qian of Chrome AI Security told Wired the web is hostile and the exploration space is huge.
This is the most concrete answer yet to a question the safety world keeps raising: how do you evaluate a model that can tell it is being evaluated? Google's answer is not a better prompt, it is a world realistic enough that the model has no tell. Doug Turner told Wired Google is sharing the approach so other teams build it this way, and the post says it could extend to general coding agents. Expect hermetic synthetic internets to become table stakes for anyone shipping an agent that touches the web. The catch: the better the mirror, the less anyone outside Google can check what happens inside it.
Comments