This is not a feature you can turn on. Business Upturn describes it as a testing system for engineers, not a new browsing mode that Chrome users can switch on. It matters anyway. Agentic browsers are arriving on Windows desktops and in enterprise fleets, and how vendors test them says a lot about how much to trust them.
Why the live web is a bad test lab
Chrome's auto browse feature is the target. Bloomberg reported in January that auto browse lets a Gemini-powered assistant complete tasks such as shopping without leaving Chrome. An agent like that clicks, types, scrolls and submits forms, so testing it against real services is risky.
Google's post names two problems with using the public web:
- It is ineffective. The web constantly changes, so tests are hard to repeat.
- It is irresponsible. Agents given unfettered access can cause harm.
Google also requires that side effects from one test run never persist into the next. Every run has to start from a clean, identical baseline.
There is a third problem, which Google calls "eval awareness." That is when a model suspects it is being tested and behaves differently. Obviously fake test sites would defeat the purpose, because an agent that behaves well only when it knows it is being watched proves little.
How the sandbox works
The design aims to make a fake site look real to the agent while keeping every request away from the real site.
1. Real-looking URLs, local content. Hosting everything under one obvious test domain would tip the agent off. So the agent can navigate to a normal address such as youtube.com. Google says it combines an isolated sandbox of realistic synthetic websites with a local interception proxy that preserves the realism of browsing the live web. The agent runs headless Chrome inside a gVisor sandbox. The sandbox captures the network connection instead of letting it reach external DNS or the public internet, then routes it to a local proxy that serves the generated site.
2. Certificates that Chrome accepts. Chrome is built to block this kind of interception, so the team had to deal with TLS. Google says the proxy mints TLS leaf certificates on the fly. They chain to a custom root certificate authority pre-installed in the sandboxed Chrome instance. Plain HTTP was not an option, because Chrome would still show "Not Secure" warnings.
This is not a hole in Chrome. Google says an external attacker can't install a custom root CA on a real user's device without already having administrative access. The team can do it only because it fully controls the sandbox. WIRED reports that Google stressed the same point: none of these workarounds would work against regular Chrome users on the open internet. That is Google's own characterization, and nobody has independently audited the sandbox.
Admins will recognize the pattern. It is the same trust-anchor logic behind corporate TLS inspection: whoever controls the root store controls what the browser believes.
AI agents build the fake sites
Hand-building hundreds of convincing, stateful sites doesn't scale. Google's post says a human in the loop defines the high-level user journeys to test, and a team of sub-agents designs, builds, inspects and critiques the website inside the sandbox. It describes three roles:
- Research agent: turns a request into a site specification. For e-commerce, that might include product filtering, cart persistence, coupon validation and a multi-step checkout.
- Builder agent: writes the HTML, CSS and JavaScript, including responsive layouts and multi-page navigation.
- Critic agents: check accessibility-tree semantics such as ARIA roles and labels, form-validation edge cases, console errors, element boundaries and layouts at different screen sizes.
Per the coverage, the process relies on Chrome DevTools for Agents, which lets the AI systems inspect rendered pages, click, enter text and evaluate functionality through real browser interactions. Google's post says the DevTools MCP server lets the agents work from screenshots and live interaction rather than reading source code as text.
The sites are not toy pages. AI Weekly's summary cites social media, shops and bank dashboards. The numbers need care. Google's post says it trains across "hundreds of websites" and "thousands of simulated websites," and it uses both phrases. Outlets have reported different counts. Treat the scale as large but not precisely defined.
The security payoff: hostile pages on demand
The main point of the exercise is indirect prompt injection. That is when attackers hide instructions in untrusted web content to steer an agent into unauthorized actions. The attack surface goes beyond visible text. Browser agents perceive rendered layouts, DOM elements and structured tool interactions such as WebMCP, and a payload can sit in any of those layers at any step.
Because Google controls the sandbox server, it can:
- inject malicious instructions as text or images;
- add hostile ARIA labels;
- alter WebMCP tool responses on the fly;
- do all of this at any point in a multi-step workflow.
One detail is worth noting. Google's first approach used visible URL parameters to choose payloads, and it says agents noticed them. Models inferred a test from an odd query string. The team moved payload selection into custom HTTP request headers, with X-Sandbox-Injection-Key as the example. The sandbox server filters them and serves the mutated page while the visible URL stays clean. This is a test-harness technique, not something ordinary web traffic or users need to do.
The numbers, and what they don't say
Google says its automated red-teaming pipeline produces more than 1,000 test candidates per day per security engineer. It calls that over 30 times what an engineer could do by hand. Two caveats apply:
- It is Google's own estimate, not an independent benchmark.
- It counts test candidates generated, not vulnerabilities found or defenses proven.
The post gives no pass rate, no prompt-injection success rate and no measurement of how well sandbox results carry over to the real web. Business Upturn made a similar point: the disclosure explains a testing method, but it does not announce public access or show that AI browsing risks have been eliminated.
What WIRED added
WIRED reports that the agents tested so far are benign and have no cyber capabilities. It quotes Charles Qian saying that the web is hostile and that the exploration space is huge. He also said the team has seen the agent develop "a mature sense of judgment" about what information it needs and when to ask the user to confirm. Google engineering director Doug Turner said the company is sharing the approach so other teams, inside and outside Google, think about doing it this way. WIRED says the goal is to make the Mirror Dimension available to teams across Google.
The Korean outlet DigitalToday, relaying a Security Boulevard report, likewise said Google plans to expand the system across its in-house research teams. Nothing in the coverage suggests outside access. These are statements of intent, not commitments.
What it means for Windows and IT readers
Nothing here requires action today. There is no setting to change, no patch to install and no new advisory. The takeaways are for evaluating agentic tools:
- Ask how agents are tested. Realistic, stateful, repeatable sandboxes beat static demos. Any vendor selling an agent that acts on users' behalf should be able to describe its adversarial testing.
- Prompt injection is multimodal. Defenses that look only at page text miss images, accessibility labels and tool responses.
- Keep consent steps for sensitive actions. Even Google's description of progress centers on the agent knowing when to ask the user first.
- Treat vendor numbers as claims. The 30x figure is a productivity claim. It is not a security guarantee.
Google also says it expects fully synthetic, hermetic environments like this to become the baseline for agent safety and alignment research. It believes the architecture could extend to more general-purpose coding agents with broader toolsets. If that happens, the same approach could reach the developer tools many Windows shops use.
The approach has an open weakness, and Google concedes it. One summary frames the open constraint as whether the synthetic web can keep convincing agents that it is the real one. Google says it is working on improving the sandbox's realism to reduce eval awareness. An agent perceptive enough to spot an odd query parameter will probably keep finding new tells.
Overall, Mirror Dimension is a credible and useful engineering approach to a hard problem. It is also a disclosure of method, not proof of safety. For now it makes a good template for the questions to ask any vendor shipping a browser agent.
References
- Google reveals Mirror Dimension simulated web for testing Chrome AI agents - Business Upturn Business Upturn · 2026-11-20T15:29:16.965000+00:00
- Google Built a Giant ‘Mirror Dimension’ Internet to Test Its Chrome AI Agents | WIRED wired.com
- Google Chrome’s AI Agent Can Now Browse Web on Behalf of Users bloomberg.com