Google Builds Synthetic Internet to Test AI Agent Security

· Original article ↗

Summary

Google says its Mirror Dimension uses simulated websites and a local interception proxy to safely test browser-use agents against malicious instructions, supporting repeated, realistic security evaluations.

Key points

  • Google researchers built the Mirror Dimension, an isolated sandbox containing thousands of simulated websites inspired by everyday user journeys.
  • A local interception proxy routes agents requesting real URLs, such as www[.]youtube[.]com, to synthetic sites without sending requests to external DNS.
  • Researchers plant malicious instructions in visible text, images, underlying page code, and screen-reader descriptions to test whether agents follow harmful directions.
  • Google says an automated red-teaming pipeline uses the environment to refine and evaluate tests, reducing reliance on hand-crafted evaluations.
  • Browser-use agents may access accounts or download information while handling sensitive user data, creating risks from scams and malicious websites.

Article Details

Event Type
Google announced a simulated internet environment for testing browser-use agents and automated red-teaming evaluations.
Impact
The isolated environment uses simulated websites and a local interception proxy so researchers can test whether agents follow malicious instructions embedded in webpage content without exposing them to real-world sites. The work addresses evaluation realism and helps researchers assess risks such as agents being tricked while accessing websites or user accounts. No real-world breach or compromise is reported.

Vendors

Products

Related Articles