🇺🇸 English 🇰🇷 한국어 🇯🇵 日本語
Home › Developer Tools
Developer Tools

Google Tests Chrome AI Agents on a Fake Web

Google's Chrome AI Security team on Nov. 19 detailed the "mirror dimension," a sealed synthetic web where auto browse agents visit real-looking addresses served by AI-built copies, letting Google run more than 1,000 automated attack tests a day per security engineer.

Google's Chrome AI Security team on Thursday, Nov. 19, detailed the "mirror dimension," a sealed-off synthetic version of the web it uses to test auto browse, the agentic feature that lets Chrome click through sites and complete tasks for users. The environment lets agents visit familiar addresses while every request is quietly served by Google-built copies, according to a Google blog post by software engineers Aaron Selya and Charles Qian.

A miniature shopping cart on a desk next to an open laptop
Shopping checkouts are among the everyday tasks Google rebuilt inside the mirror dimension. File photo · 사진: Shixart1985 / 위키미디어 공용(CC BY 2.0)

Key takeaways

  • Google runs Chrome's browsing agents in an isolated sandbox filled with synthetic shops, social feeds, recipe sites and service marketplaces instead of the live web.
  • A team of AI sub-agents builds and checks the fake sites, using Chrome DevTools for Agents to click and inspect them like a user would.
  • Google says automated red-teaming in the sandbox produces more than 1,000 attack test candidates a day per security engineer, over 30 times the manual rate.

Why test AI browser agents on a fake web?

Training and evaluating agents takes thousands of read and write actions across hundreds of websites, the engineers wrote. Pointing an agent at the public web is unreliable because pages change constantly, and risky because an agent with open access can cause real harm. Each test run also has to start from the same clean state.

The harder problem is what Google calls eval awareness: a model that suspects it is being tested may behave differently. Hosting every fake site under one obvious sandbox domain would give the game away, so the team kept real addresses such as www.youtube.com in front of the agent.

How the mirror dimension works

The agent drives a headless Chrome browser inside a gVisor sandbox. Network requests never reach external DNS or the public internet; the sandbox captures them and hands them to a local proxy that serves a generated site instead.

Because Chrome is built to block exactly this kind of interception, the proxy mints TLS certificates on the fly that chain to a custom root certificate authority preinstalled in the sandboxed browser. Google notes an outside attacker could not do this on a real user's device without administrator access.

Google says the testing ground holds thousands of simulated websites modeled on everyday user journeys. A human defines the journeys, then a research agent writes site specifications, a builder agent writes the HTML, CSS and JavaScript, and critic agents check accessibility labels, form validation, console errors and layouts. The sub-agents work through Chrome DevTools for Agents, Google's open-source MCP server, which lets them view screenshots and perform real clicks and keystrokes.

Red-teaming against prompt injection

Owning every layer lets Google plant indirect prompt injections, hidden instructions meant to hijack an agent, in page text, images, ARIA labels or altered WebMCP tool responses at any step of a multi-step task. An early version passed payload IDs in URL parameters, but frontier models noticed the odd parameters. The team now passes them in HTTP headers, so the visible URL stays clean.

A security engineer can hand-craft and test a few dozen realistic attacks a day, the post says. The automated pipeline turns each engineer's ideas into more than 1,000 test candidates daily. The post does not report how often auto browse resisted those attacks.

Context: agents move into the browser

The disclosure lands as agent features spread across Google's products; the same day, the company added computer use to its Gemini app for Mac. Researchers told WIRED the browsing agents they tested are benign and lack cyber capabilities, and Google director of engineering Doug Turner said the company aims to make the mirror dimension available to research teams across Google.

What's next

Google expects sealed sandboxes with fully synthetic internets to become the baseline for agent safety and alignment research. It says the same architecture could be extended to evaluate general-purpose coding agents with broader toolsets. Google has not announced a public release.

FAQ

Can developers use the mirror dimension? No. It is an internal Google testing environment, though Chrome DevTools for Agents, which its site-building agents rely on, is open source under the Apache 2.0 license.

Does the mirror dimension make auto browse safe? Not by itself. It is a testing tool, and Google has not published pass rates for its agents against the attacks it generates.

What is eval awareness? It is when an AI model detects that it is being tested and changes its behavior, which can make test results misleading.

References

#Google Chrome #mirror dimension #auto browse #AI agents #browser agents #prompt injection #red teaming #eval awareness #Chrome DevTools for Agents #AI safety
S
Seung Jung
AI Newsway · Developer

Editor responsible for AI Newsway. Writes about AI, developer tools, cloud and productivity technology.

More by this reporter ›
Copyright © AI Newsway All rights reserved. Unauthorized reproduction, redistribution, and AI training prohibited.
Share
f X K

Comments 0

Comments are currently disabled.

Be the first to comment.