Chrome Builds a Synthetic Web to Safely Test AI Browsing Agents

Chrome Builds a Synthetic Web to Safely Test AI Browsing Agents

A team of engineers at Chrome has created a synthetic replica of the web, also known as a mirror dimension, to safely pilot AI browsing agents. The mirror dimension uses a local interception proxy with hundreds of auto browsing websites to create a pipeline that produces more than 1000 security test candidates each day per engineer.

A key obstacle for training autonomous web agents on the live internet was intractable safety hazards and unpredictable results. We solved this by confining Chrome headless browsers within a gVisor sandbox. When an agent tries to open a real address like YouTube, the sandbox intercepts the request and forwards it to a local proxy, which serves a fictitious website. Leaving identifiable sandbox domains in the address bar reveals the model is being evaluated to sophisticated models, which degrades its training. To enable the illusion, the proxy generates TLS certs on the fly that link back to a custom Root CA installed on the sandboxed browser, which Chrome checks without security warnings.

Chrome Builds a Synthetic Web to Safely Test AI Browsing Agents

Static pages reveal themselves to leading edge models easily during testing. But manual coding lagged behind. So the team built a self serve site builder powered by three AI agents. First, a research agent charts user paths and creates low fi interface mockups with Nano Banana; then a builder agent generates HTML, CSS, and JavaScript for online shopping carts and social feeds; and finally, a critic agent examines the code through Chrome DevTools for Agents, reviewing live screenshots, clicking buttons, scrolling console logs for JS bugs, and testing accessibility tags until the simulated destination is indistinguishable from the real thing.

From a security perspective, testing for malicious website triggers and payloads is hard if possible, mostly when indirect prompt injections are embedded within an image, page text, or accessibility labels or via responses provided by the WebMCP tool. The previous version of the stress tests embedded attack triggers in plainly visible URL parameters. However, an observant AI was quick to identify the peculiar URL structure that was being used and to blow the test. The new version encodes attack instructions as part of custom HTTP request headers, which allows the server software itself to embed hidden payloads within the URL. The automation allows us to generate over 30 times the volume of attacks we could through manual effort, over 1,000 attacks for each security engineer each day.

About the author

mgtid
mgtid
Owner of Technetbook | 10+ Years of Expertise in Technology | Seasoned Writer, Designer, and Programmer | Specialist in In-Depth Tech Reviews and Industry Insights | Passionate about Driving Innovation and Educating the Tech Community Technetbook

Join the conversation

Newsletter Subscription