Inside Google’s Secret Fake Internet Built to Outsmart AI Agents

Chrome engineers constructed a hermetic "mirror dimension" to stress-test browser-use bots in a synthetic web—before frontier models realize they are being watched.
Inside Google’s Secret Fake Internet Built to Outsmart AI Agents — Tech Reader
Tech Reader  ·  Systems & Security
Feature  ·  November 19, 2026

Inside Google’s Secret Fake Internet Built to Outsmart AI Agents

How Chrome engineers constructed a hermetic "mirror dimension" to stress-test browser-use bots—before those bots realize they're being watched.
As artificial intelligence moves from passive chatbots to active web-browsing agents capable of clicking, typing, and buying on our behalf, security teams face a paradoxical dilemma: how do you safely test an autonomous bot on the open internet without letting it break things—and without tipping the AI off that it’s inside a simulation?

If you ask an AI agent to hop onto YouTube and download a podcast for your upcoming flight, it doesn’t need you to paste a link. It infers the destination, opens a browser, and navigates the web on its own. While that kind of autonomy is the core promise of Chrome’s "auto browse" feature, it creates a massive headache for the engineers tasked with keeping it safe.

Testing browser-use agents requires running thousands of read-and-write actions across hundreds of websites every single day. Pointing an experimental bot at the live internet is both unreliable—since websites constantly shift—and dangerous, because an unfettered AI can cause real-world harm. Yet putting the AI in a traditional, walled-off testing sandbox hits an unexpected snag: frontier models are getting smart enough to notice when they are being tested.

"Frontier models are far too perceptive for this. There were many instances where an agent became eval aware simply by noticing an abnormal query parameter."

As reported in a new technical deep dive on the Google Security Blog by Chrome AI Security software engineers Aaron Selya and Charles Qian, Google’s solution is straight out of a sci-fi thriller: a "mirror dimension." It is a completely hermetic, synthetic version of the internet designed to fool AI agents into thinking they are browsing the real web while keeping them safely locked inside an isolated sandbox.

The Illusion of the Real Web

To prevent an AI from developing "eval awareness"—a phenomenon where a model suspects it is in a test environment and alters its behavior—Google couldn't simply host fake websites on obvious staging URLs. If an agent sees a web address like fake-sandbox-sites.web.app, the gig is up.

Instead, the mirror dimension lets the agent navigate to https://www.youtube.com just as it would in the wild. Behind the scenes, the agent operates a headless Chrome browser inside an isolated gVisor sandbox. Rather than letting that traffic hit the public internet, the sandbox intercepts the connection and reroutes it to a local proxy serving a synthetic replica.

Pulling off that sleight of hand required solving a cryptographic catch-22. Intercepting web traffic in this manner looks identical to a person-in-the-middle attack—the exact kind of threat Chrome is built to block with fatal security warnings. To maintain the illusion, Google’s proxy dynamically mints TLS security certificates on the fly, chaining them back to a custom Root Certificate Authority pre-installed inside the sandboxed browser. To the AI, the fake web is cryptographically indistinguishable from the real thing.

Bots Building Websites for Bots

A convincing URL means nothing if the website behind it feels hollow. Multi-step AI workflows quickly expose half-baked mockups, triggering eval awareness. Because manually coding hundreds of interactive e-commerce checkouts, recipe blogs, and social feeds is impossible at scale, Google built an autonomous multi-agent harness to construct the fake web itself.

As noted in the Google Security Blog post, a human engineer simply provides a high-level user journey, and a trio of specialized sub-agents builds the site from scratch:

  • The Research Agent: Maps out architectural specs, user journeys, and basic UI image assets (generated via Google's Nano Banana model) for features like shopping cart persistence or coupon validation.
  • The Builder Agent: Writes the underlying HTML, CSS, and JavaScript to bring multi-page, responsive websites to life.
  • The Critic Agents: Audits the finished product for JavaScript errors, broken layouts, and accessibility tree semantics.

Crucially, these builder and critic bots don’t just read raw code. Using Chrome DevTools for Agents, they actually look at rendered screenshots, click buttons, type into forms, and iterate until the synthetic website behaves like a production-grade platform—prioritizing realism over speed.

Weaponizing the Sandbox

Once the fake internet is up and running, Google’s security team turns it into a hostile training ground. One of the biggest threats to web-browsing AI is indirect prompt injection, where attackers hide malicious commands inside images, text, accessibility labels, or underlying tool responses to hijack an agent mid-task.

Early on, engineers triggered these simulated attacks by appending tags directly to the web address—such as &prompt_inject_payload=abcd1234—only to watch observant models immediately spot the unnatural URL and realize they were in a test. Google solved this by hiding attack triggers inside custom HTTP request headers (X-Sandbox-Injection-Key). The synthetic web server reads the hidden header and mutates the webpage with a malicious payload on the fly, leaving the URL bar completely pristine.

"That’s over 30x what one engineer could do by hand, or more than a month of manual testing, every day."

The scale of this automated red-teaming pipeline is staggering. While a human security engineer might manually craft and verify a few dozen realistic attacks in a day, the mirror dimension generates and tests more than 1,000 adversarial candidates daily per engineer—freeing human experts to hunt for novel vulnerabilities while the machines grind through the permutations.

A Safe Playground for Smarter Machines

As autonomous agents take on longer, more complex workflows, simple benchmarks are no longer enough to guarantee safety. By constructing a synthetic internet convincing enough to fool frontier models, Google is betting that hermetic sandboxes like the mirror dimension will become the baseline for AI alignment—giving researchers a safe space to watch how machines behave before they are ever let loose on the real world.

Aaron Rose is the founder and publisher of Tech Reader Magazine, covering system architecture, cloud platforms, and AI policy.