AI News/AI News

Google trains Chrome AI agents in a simulated internet it fully controls

Google built a sealed copy of the web to test Gemini-powered browsing agents, a approach that matters to anyone whose sites agents will soon visit.

Key points

  • Google's Chrome team built the "Mirror Dimension," an isolated, fully controlled web ecosystem, to train and test Gemini-based browsing agents.
  • It includes social media, shopping and bank dashboard sites, with hidden prompt-injection attacks used to test agent defenses.
  • Website-building AI agents created the environment, and the team forged web certificates so agents could run in real Chrome.
  • Google plans to open it to more teams internally and encourages outside teams to adopt a similar approach.

Why it matters for marketers: Browser agents will act on marketers' sites and funnels, so how they are trained on checkouts, promo codes and injection attacks affects how safely they handle your pages.

Google's Chrome team has built what it calls the Mirror Dimension, an isolated web ecosystem it entirely controls, to train and test browsing agents powered by Gemini models. Google is describing the project in a blog post, according to Wired.

The environment includes social media, shopping and bank dashboard sites built with the components of real ones. Agents must complete tasks in Chrome as they would on the open web, while facing tests such as hidden prompt-injection attacks. The team says limited test setups are less effective, models can behave differently when they detect a test, and releasing unrefined agents on the real internet is risky.

To populate it quickly, the team used many website and web app development agents guided by a purpose-built harness, creating details like promo code entry and filled-in social pages. So agents could use real Chrome, researchers had to work around the browser's protections, including forging web certificates. The team stresses that none of this would work against regular users, since it relied on control over the whole network.

For marketers and content teams, the takeaway is that agents will increasingly browse, shop and fill forms on public sites, and how they are trained shapes their behavior there. Charles Qian, an engineer on Chrome AI security, said the agent is developing judgment about what information a task needs and when to ask the user to confirm.

Next, Google wants to offer the Mirror Dimension to teams across the company and hopes outside groups adopt the approach. Doug Turner, director of engineering, said investment will continue and the environment will need to grow more realistic as models improve.

Sources

This brief is our own summary. Read the full reporting at the source.

You read about AI automation. We run it.

DocFluence, built by Z Media, writes, designs, films and publishes for your brand every day, in your voice, from one login.