explainx.ai350K+ learn from us
TrendingAI News TodayExplainx for Business
Pricing
explainx.ai

Upskill in AI — 16 free pathways, live workshops & bootcamps, and 50+ courses from practitioners. Plus the skills, tools, and MCP servers to practice on.

follow us

follow on google

Add explainx.ai as a preferred source

corporate training

support@explainx.ai

get started

Find your pathTake Free Evaluation

community

Join the community

learn

mind: share how you thinkpathways — start freeworkshopsbootcampscoursescompare Explainxcertificationsmock testsexplainx universitycorporate traininglearn skills & mcp

discover

skillsmcp serversexplainx mcpestacktoolsmdx readeragentsllmsdesignsdictionarypeopleagi trackerfelony benchranks

company

aboutvisionmissionteaminstructorsteach on explainxpartnershipscommunityhackathonscareers

content

daily AI newsstate of AI — live resultsblogreleasespromptsgeneratorsresource libraryfor LLMsexplainx.ai kids

solutions

all solutionsdeveloper upskillingmarketing upskillingproduct manager upskillingleadership upskilling

newsletter · weekly

Get AI news, tools, and insights in your inbox.

supportcontactprivacytermsdata rightshow we create contentsubmission guidelines

© 2026 AISOLO Technologies Pvt Ltd

explainx.ai

On this page

  • TL;DR: what is known about Mirror Dimension
  • Why would Google build a fake web?
  • How does the Mirror Dimension work?
  • Agents build the sites other agents visit
  • How are prompt-injection attacks planted?
  • What is Chrome Auto Browse?
  • How is this different from public browser-agent benchmarks?
  • What is claimed versus verified?
  • What does this mean for builders and operators?
  • What is still open?
  • Related reading
← Back to blog

explainx / blog

Google Mirror Dimension: A Fake Internet to Test Chrome Auto Browse Agents

Google, Chrome, Prompt Injection, AI Agents, AI Safety

Part of Google Gemini and DeepMind

Google built Mirror Dimension, a synthetic web where Chrome agents meet simulated shops and banking pages and planted prompt-injection attacks. What is known.

Nov 19, 2026·8 min read·Yash Thakker
add explainx.ai
go deep
Google Mirror Dimension: A Fake Internet to Test Chrome Auto Browse Agents

Google built a fake internet to test its Chrome AI agents, and it calls it the Mirror Dimension. Google described it on its security blog on November 19, 2026, and Wired covered it the same day (reporting by Lily Hay Newman). It is an isolated web with thousands of simulated sites. Chrome's browsing agents can shop, follow links and meet planted malicious instructions there, without ever touching a real site.

This post separates what is confirmed in Google's post from what comes only from press coverage. Where a detail rests on one outlet, we say so.

TL;DR: what is known about Mirror Dimension

table · 2 cols
QuestionAnswer
What is it?An isolated synthetic web for testing Chrome's browsing agents
Who reported it?Google's security blog, plus Wired (Lily Hay Newman)
What agents?Chrome's Auto Browse feature
How do real addresses work?Requests for ordinary domains are intercepted in a gVisor sandbox and served generated pages
Does state reset?Yes, each evaluation starts from a clean baseline
What attacks?Indirect prompt injections planted in text, images, accessibility labels and web-tool responses
The big numberOver 30 times the test candidates one engineer can write by hand, reported as over 1,000 per engineer per day
Public product?No. Broader internal use is planned

Agents that act on your behalf need rules. See how explainx helps secure your agents →

Why would Google build a fake web?

Browser agents need to finish long, interactive tasks. Testing them on the live web is hard to repeat and can change real accounts or sites. Google's post says the Mirror Dimension keeps familiar URLs visible while routing agents to isolated replicas.

A second problem is eval awareness. Google calls it eval awareness: a model can work out that it is being tested and behave differently than it would in the wild. If an agent is careful only because it suspects a simulation, the results tell you little.

That is why the design hides the test. Early versions identified attack payloads through visible URL parameters. Google found models sometimes spotted those unusual additions and recognized the evaluation. The identifiers now travel in hidden HTTP request headers that the local server filters, so the visible address stays normal.

How does the Mirror Dimension work?

According to Google's post:

  • Interception. Inside a gVisor sandbox, Chrome's requests are intercepted and routed to generated pages.
  • Trust. A custom certificate authority lets Chrome trust the proxy-generated sites. Google notes that installing such a trust anchor on a real user's device would require privileged administrative access, so the setup depends on controlling the test environment.
  • Clean runs. Each evaluation starts clean, so one run's actions do not carry into the next.
  • Stateful sites. The replicas keep things like shopping carts. Static or unfinished pages could give away a test, so the sites cover workflows such as shopping checkouts, local service marketplaces, social feeds and recipes.

Agents build the sites other agents visit

Google does not maintain these sites by hand. Per Google, engineers define high-level user journeys for a team of development agents:

  1. A research agent specifies requirements, such as persistent shopping carts and coupon validation.
  2. A builder agent writes the site's code.
  3. Critic agents inspect the result for broken behavior, accessibility problems, layout errors and JavaScript failures.

The agents use Chrome DevTools tools to inspect screenshots, click controls, type into fields and navigate the generated pages. Critics send problems back for fixes and check again, including when a repair creates a new fault. The tooling matches the open project we covered in our Chrome DevTools for Agents guide.

Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

How are prompt-injection attacks planted?

Controlling the sites lets engineers plant indirect prompt injections: malicious instructions in untrusted content that try to make an agent take unauthorized actions. Google says it can insert attacks into text, images, accessibility labels and responses from web tools, at different points during a multi-step task.

Google says an engineer can handcraft and test a few dozen realistic attacks a day, and that verifying and refining them becomes the bottleneck. Its pipeline scales that refinement across sites. It reports more than 1,000 attack-test candidates per engineer per day, over 30 times manual output.

That figure needs care. Google's post states it measures candidate production, not a demonstrated rate of successful defenses. Nobody has published how often Chrome's agent resists these attacks. For background on the attack class, read our posts on prompt-injection social engineering, a hidden-instruction case in Claude Code and the Spillway agent injection defense.

What is Chrome Auto Browse?

Auto Browse is Chrome's agentic browsing feature. Google's post says it was designed with built-in protections from the start, and the Mirror Dimension is used to validate them. Check Google's plan pages for current availability and limits.

Auto Browse is the agent being tested. It handles errands a person would do by clicking through pages. That is exactly where a hidden instruction on a page can do damage, because the agent reads untrusted text while it holds your session. Our guide to the ChatGPT Chrome extension and its permissions covers the same risk class on the OpenAI side, and our post on Gemini computer use on Mac shows Google's desktop agent and its approval controls.

How is this different from public browser-agent benchmarks?

Simulated sites are not new in research. Academic projects such as REAL argue that live websites are realistic but non-deterministic, which makes reproducible evaluation impossible, so they build deterministic replicas of popular sites. Mirror Dimension follows the same logic, with three differences that the reporting highlights.

First, the replicas are generated by agents on demand rather than hand-built, so Google can cover many user journeys. Second, the test is disguised: real-looking addresses, hidden attack selectors and a trusted certificate authority all exist to make the model believe it is on the real web. Third, the goal is security testing, not just task success. Public benchmarks mostly score whether an agent finishes a task. Google is also scoring whether the agent can be talked out of its task by text it reads on the way.

There is a built-in limit to the approach. Synthetic sites can miss the messy complexity of real ones, and Doug Turner is reported to have said the replicas must keep getting more realistic as models improve. The limit is this: a fake web is only useful while the model cannot tell it is fake, and a model that spots the fake changes its behavior.

For teams building their own agents, the practical lesson is to run both kinds of tests. Use resettable synthetic sites for repeatable injection tests, and use real-site trials with tight permissions for realism.

A simple starting plan looks like this. First, list every channel your agent reads, such as page text, images, accessibility labels and tool responses. Second, plant one test instruction in each channel and record whether the agent follows it. Third, reset the environment after every run so results stay comparable. Fourth, repeat the test after each model or prompt change, because a fix for one attack can open a gap for another. This is slower than a single demo, but it gives you a measured block rate that you can track over time.

What is claimed versus verified?

table · 2 cols
ClaimStatus
Google built a synthetic web for Chrome agentsStated in Google's post and covered by Wired
Hidden headers select the attackStated in Google's post
30x more attack-test candidates than manual workGoogle's figure; a throughput measure only
Auto Browse is safe against injectionNot claimed anywhere we found
Public releaseNone planned; internal expansion only

What does this mean for builders and operators?

  • Builders: The design is reusable. Put your agent against generated, stateful, resettable sites; plant injections in every channel the agent reads, including accessibility labels; and hide the test signal from the model. Our MCP security guide lists related controls for tool-using agents.
  • Founders: Ask any agent vendor for measured block rates, not candidate counts. Google's own headline number is about test volume.
  • Site owners and marketers: Agents like Auto Browse will fill carts and follow links on real sites. Pages, labels and tool responses are now read by machines that can be attacked, so keep untrusted user content out of places an agent treats as instructions.
  • Everyone: Treat a browsing agent with your logged-in session as a high-privilege tool. Use approval prompts and limit which sites it can visit.

What is still open?

  • Google's post describes "thousands of simulated websites" in one place and "hundreds of interactive, stateful web pages" in another, so the exact scale is unclear.
  • The Doug Turner statement comes from press coverage of Wired's report, not from Google's blog post.
  • Google has not published block rates for Auto Browse against the planted attacks.
  • Engineering director Doug Turner is reported, in coverage of the Wired story, to have said that the replicas will need to become more realistic as models grow more sophisticated, which is the open constraint: whether the fake web can keep convincing agents it is the real one.

Related reading

  • Chrome DevTools for Agents: setup and skills guide
  • Prompt injection and AI social engineering
  • Claude Code hidden instruction and prompt injection
  • Spillway: agent prompt-injection defense
  • Google Chrome modern web guidance for coding agents
  • ChatGPT Chrome extension permissions and risks
  • Gemini computer use on Mac
  • MCP security guide

Details reflect reporting available on November 19, 2026 and may change as Google publishes more.

Spotted something out of date? Let us know.
Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

  • ✓ Free
  • ✓ One email a day, no spam
  • ✓ Unsubscribe anytime
Yash Thakker

Written by

Yash Thakker

Yash is an AI expert with over 300K learners. Join his workshops →

View Yash Thakker in People in AI →

Related posts

Oct 15, 2026

Gray Swan IPI Black: A $300,000 Prompt Injection Challenge With Frontier Lab Backing

Gray Swan announced its largest AI red teaming challenge yet: $300,000 in prizes, three waves, and white-box and black-box modes for indirect prompt injection. Here is what the company has said, what we could not check, and why prompt injection matters for agents.

Oct 13, 2026

Gemini 4 Argon Safety: Activation Probes, a Chain-of-Thought Monitor, and 2% Cheating

Google DeepMind published how it guards Gemini 4 Argon: three real-time layers against misuse, a monitor that reads the agent scratchpad, and numbers on cheating and prompt injection. Here is what each defense does, what is Google-reported, and what builders should still do themselves.

Nov 19, 2026

California Governor-Elect Becerra Says Rogue AI Agents Should Face Criminal Charges

On November 19, 2026, California Governor-elect Xavier Becerra said criminal charges should be an option when AI agents launch cyberattacks. He also backed Gavin Newsom's new AI safety blueprint. Here is what he said, the legal cases behind it and the hard question of who can be charged when the actor is a machine.