Chrome Security Team Introduces Mirror Dimension for Evaluating Browser Agents in Hermetic Sandboxes
·2026.11.20 02:00
Key point
The synthetic web environment uses a gVisor-based sandbox and multi-agent harness to automate red-teaming, increasing efficiency by over 30x compared to manual testing.
Details
- Chrome Security team launched Mirror Dimension, a synthetic web environment for evaluating browser-use agents without real-world security risks.
- Automated red-teaming pipeline generates and refines over 1,000 test candidates per engineer daily, a 30x efficiency gain over manual testing which yields only dozens per day.
- Architecture uses a gVisor-based hermetic sandbox with an interception proxy that blocks external DNS and routes traffic to local synthetic sites while maintaining real URLs to prevent eval awareness.
- A multi-agent site development harness autonomously generates hundreds of interactive, stateful web pages using Research, Builder, and Critic agents supported by Chrome DevTools for Agents (MCP).
- Dynamic red-teaming injects adversarial payloads via multiple vectors (images, text, ARIA labels, WebMCP tool responses) using HTTP headers to hide injection identifiers from the agent.
- Future plans include extending the architecture to general-purpose coding agents with broader toolsets and improving sandbox realism to minimize eval awareness.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.