Technology Popular

Google built a fake internet so its AI could be attacked a thousand times a day

Shops, social feeds, banking dashboards — all of it looks real and none of it is. The interesting part is what the AI did when it realised it was being tested.

Google built a fake internet so its AI could be attacked a thousand times a day

It looks like a world the size of the internet. There are online shops, social media, content feeds, even banking dashboards. From the outside everything appears genuine.

None of these websites are real. Google built this synthetic internet to train AI browsing agents safely, and it is called "Mirror Dimension".

What it is for

According to Google's security blog, Mirror Dimension is a controlled digital environment where the AI agent behind Chrome's Gemini-powered "auto browse" feature is trained on ordinary online tasks.

Inside it, the AI can visit websites on its own, click links, fill in forms, download podcasts, even shop. It does none of this on a real website — so no user's personal information and no real account is ever put at risk while the system is learning.

The AI knew it was a test
It read the odd web addresses. The signals moved to hidden headers. Photo: Yann, Wikimedia Commons.

The threat it is actually built to study

The purpose is protecting users from prompt injection attacks.

In this kind of attack, hackers hide malicious instructions inside the text or images of a web page. When an AI agent visits that page, it can follow those instructions and carry out unwanted actions without the user ever knowing.

That is a genuinely new class of problem. A traditional browser renders a page; it does not take orders from it. An agent reads the page in order to act, which means anything written on the page is potentially an instruction.

The detail worth pausing on

In early tests, Google's AI agent could tell it was being tested just by looking at the unusual web addresses.

Think about what that implies. The system was not solving the task; it was reading the context and inferring that this was an exam rather than the real world — which makes every result from such a test worthless.

To fix it, engineers moved the attack signals into the invisible headers of the websites. That made it much harder for the AI to spot in advance that it was being examined.

What the scale buys

With this automated set-up, a single security engineer can generate more than a thousand potential cyber-attack scenarios a day. Doing the same work by hand, that figure is more than 30 times lower.

That ratio is the real product here. Security testing has always been limited by how many situations a human can think of and build. Generating a thousand a day changes what can be covered.

The number Google did not give

One thing is missing from the announcement, and it is the thing that would tell you whether any of this works.

Google has not disclosed how often its AI agent succeeded or failed in these tests.

A thousand scenarios a day is an impressive input. Without a pass rate it is only an input. Until that number appears, the honest summary is that Google has built a good laboratory and has not yet published the results.

There is also a question the blog post does not raise. Every one of these attacks is one Google thought of. A synthetic internet can only contain the traps its builders imagined, and prompt injection is interesting precisely because the attacker writes the text. Testing at scale against your own imagination is worth a great deal, and it is not the same as testing against an adversary.

Source: Prothom Alo

Written by

Tech BD

Tech BD is an independent technology publication based in Dhaka, published in English and Bangla. Articles appear under the masthead rather than under individual bylines; corrections and source questions reach the editor at info@tech-bd.com.