A web that isn’t there
Google has built a synthetic copy of the web to test the artificial intelligence agents inside Chrome, giving them thousands of fake websites to browse without ever touching the real internet. The system, called the mirror dimension, lets engineers attack their own agents with hidden instructions in a sealed environment.
Chrome’s agentic browsing feature, called auto browse, carries out multi-step tasks on behalf of users, from shopping to booking. Testing it on the live web is risky, because an agent can cause real harm, and unreliable, because the web changes constantly. Google needed a controlled ground instead.
The mirror dimension is a hermetic sandbox filled with realistic synthetic sites inspired by everyday journeys, from local marketplaces to checkouts and social feeds. The environment resets after every run, so side effects from one test never leak into the next. Every run starts from a clean baseline.
How the illusion holds
The hardest problem was keeping the agent from realising it was being tested. A fake sandbox address instantly tips off a frontier model, a problem Google calls eval awareness. So the agent navigates to real addresses such as youtube.com, while a local proxy quietly serves a generated copy instead.
The proxy runs inside a gVisor sandbox, and the agent uses a headless Chrome browser that never reaches the public internet. To stop Chrome blocking the interception as an attack, the proxy mints TLS certificates on the fly that chain to a custom root authority installed only inside the sandbox.
Teaching the web to attack itself
Google also uses the mirror dimension to stress-test its agents against indirect prompt injection, where hidden instructions inside web content try to hijack an agent. The attacks can hide in visible text, images, accessibility labels or tool responses, at any point in a multi-step task.
Building the fake sites is itself done by agents. A research agent writes the specifications, a builder agent writes the code, and critic agents check the result against production benchmarks. The harness uses Chrome DevTools for Agents, which lets the agents see screenshots and click through the pages they build.
Google first tried passing attack payloads in web addresses, but frontier models noticed the strange parameters and became eval aware. The team now sends payload identifiers through HTTP headers instead, leaving the visible address pristine. To the agent, the page looks entirely normal.
Our Read
The mirror dimension is Google admitting, in public, that testing browser agents is a security problem in its own right. The striking part is the cat-and-mouse game: the testers must fool models that are trained to spot exactly this kind of deception. Eval awareness is the new benchmark gaming.
The scale is the other signal. A security engineer can hand-craft a few dozen attacks a day, but the pipeline now produces over 1,000 test candidates per engineer daily. That is over thirty times the manual rate — more than a month of hand testing in a single day.
Why It Matters
Browser agents will soon do far more for us, from shopping to banking, and every page they visit is untrusted input. If a hidden instruction can redirect an agent, the damage is real money. Google’s answer is to rehearse those attacks thousands of times before users meet them.
Background
Google’s post, published on 19 November by Chrome AI Security engineers Aaron Selya and Charles Qian, says the mirror dimension will also be extended to coding agents with broader toolsets. The company plans to keep expanding the sandbox as agents take on longer, more capable workflows.
Source: Google’s security blog


0 Comments