Google created an isolated web world "Mirror Dimension" for training and testing AI agents in Chrome
Google has built a closed, fully controlled web ecosystem called "Mirror Dimension" with copies of e-shops, social networks, and banking dashboards, where it trains and security-tests web-use agents in Chrome, including resilience against prompt-injection attacks.
Google has for the first time publicly described the "Mirror Dimension" project – a large-scale but isolated web environment fully controlled by the Chrome team. It contains a wide range of realistic websites and platforms (e-shops, social networks, banking dashboards) with functionality matching real sites, such as entering promo codes or populated social media profiles. In this environment, browsing assistant agents perform tasks just as they would on the real internet and undergo resilience testing, including hidden prompt-injection attacks.
According to Google, the reason for creating its own environment is that agents need experience with the open internet, but letting immature agents loose on the real web would be risky; at the same time, models can exhibit so-called eval awareness, behaving differently if they recognize a test environment. Mirror Dimension was built by the Chrome team deploying a large number of agents for developing websites and web applications, controlled by a purpose-built AI harness, to quickly create realistic web services and data flows at sufficient scale and diversity.
For the agents to be able to use the real Chrome browser within the environment (which the researchers say is key for training accuracy), the team had to bypass a number of Chrome's protective mechanisms and verification checks, including creating fake web certificates. The company emphasizes that these procedures would not be possible against regular users on the open internet – they work only because Google has full control over the entire environment and network. According to the company, the agents being tested are harmless and have no cyber capabilities.
Google plans to make Mirror Dimension available to other teams within the company, and according to engineer Doug Turner, this is the first step in a long-term investment in this approach, with further expansion of the environment's realism expected as agents grow more sophisticated. The company is also sharing this approach externally so that other organizations might consider a similar solution.
Why it matters
Testing AI agents on the real internet carries the risk of harming actual services and users; an isolated but realistic copy of the web makes it possible to safely verify how an agent responds to prompt-injection-type attacks without the threat of real damage. For companies developing their own AI agents, this is a concrete illustration of a methodology for building such testing – including the fact that assembling a realistic test environment itself is a non-trivial task, one that Google addressed by deploying additional AI agents to build websites.
Two audiences, two different impacts
What this means
For individuals
For the average Chrome user, this means agentic features in the browser should go through more thorough security testing before deployment, which could reduce the risk of successful prompt-injection attacks when using an AI assistant in the browser.
More practical updates →For a business
Companies developing or deploying web-use AI agents gain documented inspiration for their own security testing – building an isolated but realistic test environment instead of relying on limited sandboxes or risky tests on the live internet.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.