Google built a copy of the internet for AI agents and sped up attack testing more than 30-fold
Google has built Mirror Dimension, a closed copy of the internet with fake stores, social networks and banks, so it can test AI agents without putting real accounts at risk. An agent sees the YouTube address, but lands on a replica where hidden fraudulent instructions are automatically planted. A security engineer can prepare more than 1,000 test attacks per day, over 30 times the manual rate. For browser agents, this sandbox is the more practical place to run these checks.
AI-processed from RB.RU; edited by Hamidun News
Imagine asking an AI agent to find the latest episode of a podcast on YouTube and prepare it for download before your flight. The agent decides for itself where to go, what to click and which forms to fill in. Now imagine that it made a mistake, or that it was tricked by hidden text on a page, and all of this happened in a real account.
To avoid testing agents on live people and companies, Google built Mirror Dimension, an isolated copy of the internet. According to the company, one security engineer can create more than 1,000 test attack variants per day inside it, which is more than 30 times what a specialist can prepare and check by hand.
What Mirror Dimension is
According to Google, the environment reproduces thousands of plausible sites and services: social networks, online stores, banking apps and classifieds platforms. The developers copied more than just the appearance. The pages have to work: the agent can follow links, fill in forms, press buttons and complete multi-step tasks.
Why build it? Google cites two reasons why training and testing agents on the open internet is impractical:
- Page content changes constantly, so test results cannot be reliably compared between runs.
- The agent's actions can affect real accounts, users and companies.
In addition, users rarely name the exact address of a site. The agent has to work out for itself which resource is needed, open it and carry out a chain of actions. Google estimates that properly testing this behavior requires thousands of runs across hundreds of sites. On live services, that is expensive and risky.
How the agent is convinced the internet is real
Technically, the setup works like this: a local server intercepts the browser's requests and routes them to synthetic sites. The agent sees familiar addresses, such as YouTube, but is in fact working with an isolated copy.
The developers named the main challenge as whether the agent will realize it is being tested. If it notices signs of a simulation, it may behave differently than it would in real life, and the test results will be less reliable. For that reason, Google uses its own fake security certificates. The browser treats the connection to the local proxy as trusted, so the agent has no grounds for suspicion.
Developers will recognize the logic. When you test a payment module, you do not charge real cards; you use a sandbox. But agents need a much broader sandbox: they require not individual APIs but an entire "city" of sites in which they can move freely.
Testing for deception through page content
The second purpose of the environment is security. Sites can hide instructions that make an agent forget the user's task and do something unrelated. For example, text on a page might demand that data be sent or that a malicious link be opened. Mirror Dimension automatically inserts such fraudulent commands into various page elements, and engineers observe how the model responds and correct its behavior.
Google's figures are as follows: the environment allows more than 1,000 test attack variants per day per security engineer. That is more than 30 times the number of scenarios a person can prepare and verify manually. An important caveat: these are the company's own data, and the article contains no independent verification. There is also no information on how accurately the copies reproduce complex sites or how often the agent still recognizes the substitution.
What this means in practice
If you build browser agents, the takeaway is simple: testing them only on the live internet is no longer enough. You need a reproducible environment where pages do not change on their own and where failure can be staged safely. You also need an automatic attack generator, because writing that many scenarios by hand is unrealistic.
For product managers there is another implication. The more users hand agents purchases, transfers and account management, the more a single successful attack through hidden text on a page will cost. Companies that run agents through thousands of such scenarios in advance will gain a clear advantage in trust. Mirror Dimension currently looks like an internal Google tool, but the approach itself, a synthetic internet plus automated attacks, will likely become a standard across the industry.
- Nov 21, 2026A swimming team captain handed ChatGPT three scheduled tasks and cut her routine work by a multiple
- Nov 21, 2026MIRA Neural Network Renders Rocket League at 20 FPS, but Ping Hits 460 ms
- Nov 21, 2026Google built a closed copy of the internet with fake banks to train Chrome's AI agents
- Nov 21, 2026MLPerf Training Adds Nvidia Rubin and TPU Ironwood in First Appearance: 153 Systems
Key Questions Answered
What Mirror Dimension is?
According to Google, the environment reproduces thousands of plausible sites and services: social networks, online stores, banking apps and classifieds platforms. The developers copied more than just the appearance. The pages have to work: the agent can follow links, fill in forms, press buttons and complete multi-step tasks. Why build it?
How the agent is convinced the internet is real?
Technically, the setup works like this: a local server intercepts the browser's requests and routes them to synthetic sites. The agent sees familiar addresses, such as YouTube, but is in fact working with an isolated copy. The developers named the main challenge as whether the agent will realize it is being tested. If it notices signs of a simulation, it may behave differently than it would in real life, and the test results will be less reliable. For that reason, Google uses its own fake security certificates.
What are the key details on Testing for deception through page content?
The second purpose of the environment is security. Sites can hide instructions that make an agent forget the user's task and do something unrelated. For example, text on a page might demand that data be sent or that a malicious link be opened. Mirror Dimension automatically inserts such fraudulent commands into various page elements, and engineers observe how the model responds and correct its behavior.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.