Google builds isolated 'Mirror Dimension' web to test Chrome AI browsing agents

Google's Chrome team created a fully controlled copy of the web, with fake shops, social sites and bank dashboards, to train and stress-test Gemini-based browsing agents safely.

· 2 min read

a close up of a network with wires connected to it
Photo: Albert Stoynov / Unsplash

Google has built a sealed-off, simulated version of the internet to test the AI browsing agents it is developing for Chrome, according to a report from WIRED. The environment is called the Mirror Dimension, and Google is describing it publicly for the first time in a blog post.

The Chrome team, which is working on browsing features powered by Google's Gemini models, says web-use agents have to understand real websites and be hardened against online attacks. Limited test setups only go so far in providing that. Letting immature agents loose on the open web is also risky, so the team chose to build a web it fully controls.

Another concern is what researchers call eval awareness. This is when a model behaves differently because it detects that it is in a test or simulation. A broad and realistic environment is meant to reduce that problem.

The Mirror Dimension contains a wide range of sites, including social media platforms, shopping sites and bank dashboards, built with the components and attributes of real websites. Agents must complete tasks in Chrome as they would on the actual internet. Along the way they face resilience tests, such as hidden prompt-injection attacks designed to probe their defenses.

Charles Qian, a software engineer on the Chrome AI security team, told WIRED that agents are not programmed, so every possible situation cannot be predicted. He described the web as hostile and said the team must explore that space as thoroughly as it can. He added that the agent has begun to show a mature sense of judgment about what information a task needs and when to hand control back to the user for confirmation.

Populating the environment was a challenge in itself, since it had to be large and varied while still being built quickly. According to the report, the team used groups of website and web app development agents, guided by a purpose-built AI harness, to create realistic services and data flows. These include details such as entering promo codes on a shopping site or filling social media pages with content.

To let the agents run in genuine Chrome, which is important for accurate training, the researchers had to work around the browser's many user protections and validity checks. They effectively hacked the browser and took steps such as forging web certificates. The team stressed that such attacks would not work against ordinary Chrome users on the open internet, because they depend on the full access that comes with controlling the whole network.

The researchers say the agents under test are benign and have no cyber capabilities. Their goal is to make the Mirror Dimension available to teams across Google. Doug Turner, Google's director of engineering, said the company is sharing the approach so other teams inside and outside Google consider it, and expects continued investment as models grow more sophisticated and the environment needs to become more realistic.

Related

Google builds isolated 'Mirror Dimension' web to test Chrome AI browsing agents | LTCMiner News