Google’s Chrome team has built a private, simulated internet to test auto browse, the browser’s agent that clicks, scrolls, and types on a user’s behalf. Engineers call it the “mirror dimension.” According to a Chrome Security post on Google’s blog dated November 19, it holds thousands of generated websites modeled on everyday tasks like shopping, recipes, and social feeds. The point is to attack the agent thoroughly without letting it near anything real.

The design problem is that a test agent that knows it is being tested stops behaving normally. Google’s engineers call this eval awareness. Their first approach, hosting every fake site under one made-up domain, failed on realism, and telling the agent to visit a sandbox address in the prompt did the same.

So the agent is told to visit youtube.com, and it goes to youtube.com. Inside the sandbox, a headless Chrome instance runs under gVisor, and the network layer diverts that request to a proxy running locally, which answers with a generated stand-in. Chrome is built to reject exactly this kind of interception, so the proxy mints a fresh TLS certificate for each site on demand. Those certificates chain to a custom root authority installed in the sandboxed browser at startup. Google notes that a real attacker could not do this without administrator access to the machine, which is why the trick is safe here.

Populating the fake web is the second problem. Hand-built sites feel static, and the post says agents pick up on that. Google’s answer is a harness of three agent roles: a research agent drafts the specification, a builder agent writes the HTML, CSS, and JavaScript, and critic agents check accessibility semantics, form validation, console errors, and layouts at different screen sizes. The critics work through the Chrome DevTools MCP server, so they take screenshots and click through pages instead of reading source. A human supplies only the user journey to cover.

The security payoff is in the red-team half. Because Google controls the server, it can plant a prompt injection in visible text, an image, an ARIA label, or a WebMCP tool response, at any step of a multi-step task. The team first passed payload identifiers as URL parameters, then found that frontier models noticed the odd parameter and flagged the test. Payload keys now travel in HTTP headers such as X-Sandbox-Injection-Key, leaving the visible URL clean.

Google puts a number on the gain. A security engineer can hand-craft a few dozen attacks per day, while the pipeline generates more than 1,000 candidates per engineer daily, which the post describes as over 30 times the manual rate. The post gives no figures on how often auto browse resisted those attacks, and no comparison with an unprotected agent. All the evidence is Google testing its own product in an environment it built.

That gap is the part to weigh. The mirror dimension shows that a credible agent eval is now partly a deception project: every URL, header, and certificate has to look ordinary to a model that is watching for tells. Teams running agent evals against a staging domain or a flagged test account should assume the model can see the seams, and that a safety score earned there may not hold on the open web.

Google, Chrome Security blog post “Mirror Dimension for Agents: Evaluating Browser-Use Agents Without the Web,” November 19, 2026.