Google Built a Giant ‘Mirror Dimension’ Internet to Test Its Chrome AI Agents
AI-assisted digest, checked against the original source before publishing. How we work
Google's Chrome team built a fully controlled alternate internet—dubbed the 'Mirror Dimension'—to train and stress-test Gemini-powered browsing agents against prompt injections and other web attacks.
Google's Chrome team has built a sprawling, isolated copy of the internet to train the AI browsing agents it's developing on Gemini models. The company detailed the project, called the "Mirror Dimension," in a blog post for the first time. The rationale: web-use agents need exposure to real-world sites and platforms to be effective, but running unrefined agents against the live web is dangerous, and narrow test environments are too limited to be convincing.
Google also cited a subtler problem known as "eval awareness," where models behave differently once they detect they are inside a test or simulation. The Mirror Dimension is meant to sidestep both issues—a diverse, isolated web ecosystem entirely under Google's control that still mimics the structure of the open internet.
The environment includes a wide range of sites and platforms, from social media and shopping to bank browsing assistants, all carrying the components and attributes of real websites. Agents must complete tasks in Chrome just as they would on the open web, while facing resilience tests along the way, including hidden prompt-injection attacks designed to probe their defenses. Charles Qian, a software engineer on the Chrome AI security team, said the exploration space is vast because agents aren't pre-programmed: "Any user can ask any prompt on any website, so we really need to explore that space as thoroughly as we can." He added that agents have shown "a mature sense of judgement" about which information a task actually requires and when to hand a decision back to the user for confirmation.
Populating the environment posed its own engineering challenge: it needed to be large and varied enough to resemble the real web, yet built fast. The Chrome team deployed armies of website and web-app development agents, coordinated by a purpose-built AI harness, to generate realistic web services and data flows—right down to promo-code entry on shopping sites and populated social media pages. To let agents run in an unmodified Chrome browser, which the team considered crucial for accurate training, researchers essentially had to subvert the browser itself, forging web certificates to bypass validity checks. Google stresses those steps only worked because of the privileged, environment-wide access the team holds and would not be exploitable against ordinary Chrome users.
Google says the browsing assistants tested so far are benign and lack cyber capabilities, but it wants to open the Mirror Dimension to teams across Google—and to encourage outside researchers to adopt the approach. "We're sharing this approach so other teams both internally and externally think about doing it this way," said Doug Turner, Google's director of engineering, calling it the first step with continued investment to come.
The open questions: whether the synthetic web can be kept realistic enough as agents grow more capable, how much fidelity is lost when a controlled environment substitutes for the genuine messiness of the open internet, and whether the approach becomes a shared industry standard for agent training or remains an internal Google advantage.
Key facts
- • Google revealed its 'Mirror Dimension'—a fully controlled alternate internet for training Chrome AI agents.
- • Environment includes social media, shopping, and bank dashboard sites with realistic data flows.
- • Agents face hidden prompt-injection attacks during resilience testing in real Chrome.
- • Chrome team used armies of dev agents plus a custom AI harness to build sites.
- • Google plans to share the approach with internal and external researchers.
Why it matters
Teams building web agents now have a concrete template for scaling training environments without exposing the live internet to risk—and without triggering eval-awareness blind spots. Expect synthetic-web sandboxes to become a competitive axis, with realism and scale as the metrics that separate credible agent training from narrow benchmark theater.
Spotted an error? Send a correction.
Get the day's AI news in one email
The Daily Epoch · 60-second read · unsubscribe anytime