Skip to content
Advertisement
Google unveils Mirror Dimension to train Chrome AI agents safely

Google unveils Mirror Dimension to train Chrome AI agents safely

Google / Wikimedia Commons (Public domain)

Google is tackling a stubborn problem: AI agents that learn in sandboxes but have to work on the messy, hostile open web

Google has introduced a training setup called Mirror Dimension, built to prepare the AI agents inside Chrome for life on the real internet, according to Wired.

The name shares a title with a Marvel location where heroes fight without wrecking actual cities. The goal here sounds similar: let agents practice without breaking anything that matters.

Why training a web agent is harder than it looks

So-called web-use agents are AI systems that browse on your behalf. They click buttons, read pages, fill in fields and move between sites the way a person would.

To be useful, they need to understand how real websites and platforms behave. They also need to hold up against attacks hiding in those pages.

There’s a second, stranger problem. AI models can develop what researchers call eval awareness.

Put plainly, a model can figure out that it’s being tested and start behaving differently than it would in the wild. If an agent acts carefully in a simulation because it suspects a simulation, the test results stop telling you much. Mirror Dimension is Google’s attempt to close that gap between the practice field and the real game.

The agent Google is training already has a job

This isn’t a purely academic exercise. Google already has an agent working inside Chrome.

Advertisement

The company began rolling out a feature called Auto Browse on January 28, 2026. It runs on the Gemini 3 model and is available to AI Pro and AI Ultra subscribers in the US.

The limits are tiered. AI Pro users get 20 tasks per day, while AI Ultra users get 200.

The feature handles the kind of errands most people put off. It can research flights, fill out forms, manage subscriptions and even collect tax documents.

Google has built some guardrails in. Auto Browse asks users for explicit confirmation before sensitive actions like logins and purchases.

A layered approach to agent safety

Mirror Dimension fits into a broader security push that Google has been building out over the past several months.

In December 2025, Google published a post on its security architecture that introduced what it calls “user alignment critic” models. These are designed to boost safety by checking whether agent behavior lines up with what the user actually wants.

On the enterprise side, Google’s Gemini Enterprise Agent Platform offers containerized browser environments for agent automation. These sandboxes are meant to protect users and their data while agents carry out browsing tasks.

Then came Google I/O in May 2026. There, the company highlighted WebMCP, a proposed standard for giving agents structured tools, along with new guidelines for AI coding.

The idea behind a structured standard is straightforward. Instead of forcing agents to guess their way through pages built for human eyes, websites could expose clear, machine-readable tools that agents can call directly.

What this means for users, developers and the web

The biggest stakes are for everyday Chrome users. Browser agents are moving from demos to features people pay for, and the cost of a mistake rises with every task they take on.

That’s why the eval awareness issue deserves attention. If models can tell when they’re being tested, then passing a safety test is weaker evidence than it appears.

For website owners and developers, the direction of travel is also becoming clearer. Between WebMCP and in-browser agents, sites may increasingly be visited by software acting for humans rather than humans themselves.

The pricing structure hints at how Google is managing that risk. Capping Pro users at 20 tasks and Ultra users at 200 keeps usage measured while the systems mature.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.