Google has revealed how it tests the AI agent inside Chrome without letting it loose on the real web. The Chrome security team built a “mirror dimension”, a sealed-off synthetic internet with thousands of simulated websites, where the browser agent can shop, browse and fill in forms while engineers attack it with hidden instructions.
Quick Take
- What it is: a hermetic test environment for Chrome’s agentic browsing feature, auto browse, with thousands of fake sites modeled on everyday tasks.
- Same URLs: the agent visits real addresses like youtube.com, but a local proxy serves a generated copy instead of the live site.
- Built by agents: research, builder and critic agents create the test sites, using Chrome DevTools for Agents.
- Red-teaming at scale: Google says the pipeline creates more than 1,000 test candidates per security engineer per day, over 30 times what one engineer could do by hand.
In plain English
AI agents in a browser can click, type and buy things for you, so a mistake or a trick can cost real money. Testing them on the real internet is risky and the web changes all the time, so results are hard to repeat. Google built a fake copy of the web that looks real to the agent. There it can test whether the agent finishes tasks and whether hidden instructions on a page can trick it, without any harm to real people.
What happened
In a Google blog post, Chrome AI Security engineers Aaron Selya and Charles Qian describe the mirror dimension as “a synthetic version of the web to test agents safely.” It lets engineers evaluate auto browse “in realistic online environments, featuring thousands of simulated websites inspired by everyday user journeys, without creating real-world security risks.”
Wired, which reported on the system on 19 November 2026, said Google plans to make the environment available to more teams across the company.
How it works
- Sandbox and proxy: the agent runs a headless Chrome browser inside an isolated gVisor sandbox. Network requests never reach the public internet; they are routed to a local proxy that serves the generated site.
- No tell-tale addresses: Google found that fake sandbox URLs made models suspect a test, a problem called eval awareness. So the agent navigates to normal addresses, and the proxy mints TLS certificates on the fly that chain to a custom root certificate installed only inside the sandbox.
- Sites built by agents: 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 until the site works.
- Hidden attacks: the test server can inject prompt injections as text or images, in ARIA labels or in WebMCP tool responses. Payload IDs travel in HTTP headers, so the visible URL stays clean and the agent sees nothing unusual.
Why it matters
For users, this is the testing behind features like Google Auto Browse in Chrome. Prompt injection, where a web page hides instructions that hijack an agent, is one of the main risks of letting AI act in a browser.
For developers building agents, the post is a detailed blueprint: repeatable, hermetic tests with realistic sites, and adversarial payloads that can be placed anywhere in the page or tool output. Google says the same architecture could be extended to coding agents with broader toolsets.
For the industry, Google writes that it expects “hermetic sandboxes equipped with fully synthetic internets like the mirror dimension will be the baseline for agent safety and alignment research.”
What’s still open
- How well results in a synthetic web carry over to the messy real one.
- Whether Google will release the tooling or test sites to outside researchers.
- How auto browse scores on these tests; Google has not published pass rates.
Sources
- Google blog, “Mirror Dimension for Agents: Evaluating Browser-Use Agents Without the Web,” Aaron Selya and Charles Qian, Chrome AI Security
- Wired, “Google Built a Giant ‘Mirror Dimension’ Internet to Test Its Chrome AI Agents,” 19 November 2026
i10x Perspective
Browser agents will only be trusted if they can resist booby-trapped pages, and you cannot prove that by testing on the live web. A fake internet built by agents to attack another agent is a smart answer. Expect other browser makers to copy the idea. More in Browsers & Assistants.






