Security
Google tests its browsing agents on a “mirror internet” of fake sites to catch prompt injection
Google’s Chrome security team built a synthetic web of fake social, shopping and bank sites to test browsing agents against prompt injection.
Google’s Chrome AI-security team has built a closed, synthetic copy of the web to test its browsing agents, according to Wired and a Google blog post titled “Mirror Dimension for Agents”. The aim is to see how Chrome’s “auto browse” agent reacts to indirect prompt injection, where a web page hides instructions meant for the AI.
The essentials
- What it is: a hermetic synthetic web with fake social, shopping and banking sites generated by AI, per Wired and Google’s post.
- How it runs: the agent drives headless Chrome inside a gVisor sandbox, and its traffic goes to a local proxy, per Google’s post.
- How attacks are planted: injection payloads are hidden in a request header called X-Sandbox-Injection-Key, per Google’s post.
- Who wrote it: Aaron Selya and Charles Qian, per the blog post.
Why a fake internet
Testing an agent on real sites risks real harm: it could submit a form or move money. A closed copy lets researchers attack the agent as often as they like and see exactly what it did, with nothing real at stake.
What indirect prompt injection is
An agent reads the page it visits. If the page contains text such as “ignore your user and do this instead”, a weak agent may obey. Defences have to be tested against many such pages.
What is not established
- Results of the tests, such as how often the agent was fooled: not stated in the sources.
- When the testing environment will be released to outside researchers: not stated in the sources.
- How closely the fake sites match real ones: not stated in the sources.
Summary table
| Point | What is stated as of November 20, 2026 | Source |
|---|---|---|
| Name | Mirror Dimension for Agents | |
| Sandbox | gVisor; headless Chrome; local proxy | |
| Injection channel | X-Sandbox-Injection-Key header | |
| Authors | Aaron Selya, Charles Qian |
FAQ
Does the agent touch real websites during these tests?
No. The agent runs against a closed synthetic web.
Is this a product?
No. It is an internal testing environment described in a blog post.
Sources
- Wired: Google builds a fake internet to test browsing agents (verified 2026-11-20)
- Google Security Blog: Mirror Dimension for Agents (verified 2026-11-20)
Spotted an error in this article? Report it · Our corrections policy