The Daily Byte
Breaking
Welcome to The Daily Byte — real stories from across Pakistan.
Saturday, 21 November 2026 · Pakistan

The Daily Byte

Stories from Pakistan that deserve to be heard
Minor Technology

Google builds fake internet to test Chrome AI agents

By The Daily Byte ·20 Nov 2026 ·3 min read
Laptop on a desk with the Google Chrome logo displayed on its screen

Google has built a synthetic copy of the internet to test its Chrome AI browsing agents, so they can be put through their paces without touching the real web or real users.

The system, called Mirror Dimension, was described by Google's Chrome AI Security team in a blog post published on 19 November 2026. It contains thousands of simulated websites inspired by everyday journeys, from shopping checkouts and social media feeds to bank dashboards and recipe sites, where agents can browse, click, type and complete tasks as they would in the wild.

Why agents cannot be tested on the live web

Chrome's agentic browsing feature, known as auto browse, lets an AI agent carry out multi-step tasks on a user's behalf. Testing it on the public web is both unreliable, because the web constantly changes, and risky, because an agent given free rein can cause real harm. Google says rigorous testing must be safe, reliable and repeatable, with every test run starting from a clean, identical baseline.

Mirror Dimension solves that by running a headless Chrome browser inside an isolated gVisor sandbox. A local proxy intercepts the browser's network requests and serves a generated site instead, so nothing reaches external DNS or the public internet. The agent still navigates to the real address, such as youtube.com, so it has no clue it is inside a test.

Forged certificates keep the illusion intact

That interception is, in effect, the kind of person-in-the-middle attack Chrome is built to block, so the team had to work around its own security. The proxy mints TLS certificates on the fly, chaining back to a custom root certificate installed in the sandboxed Chrome, so the browser validates the synthetic sites normally. Google stresses this is only possible because it fully controls the sandbox, and that none of these techniques would work against regular Chrome users.

The synthetic sites themselves are built by a multi-agent harness. A research agent writes the specification, a builder agent writes the HTML, CSS and JavaScript, and critic agents check accessibility, form validation, JavaScript errors and responsive layouts, using Chrome DevTools for Agents to inspect screenshots and carry out clicks and keystrokes.

Stress-testing against prompt injection

The main purpose is security testing. Mirror Dimension lets Google bombard agents with indirect prompt injection, where hidden instructions embedded in web content try to manipulate an agent into unauthorised actions. Malicious payloads can be hidden in text, images, ARIA labels or WebMCP tool responses, and the test harness passes payload identifiers through a custom HTTP header so the page's visible URL stays pristine and the agent is not tipped off.

Google says the automated red-teaming pipeline now generates more than 1,000 attack-test candidates per security engineer per day, over 30 times what one engineer could produce by hand. The company expects hermetic sandboxes with fully synthetic internets to become the baseline for agent safety research as browser agents take on longer, more complex tasks.