AI · Nov 18, 2026
Gemini comes to budget Android phones, and that is a bigger deal than it soundsGoogle’s “Mirror Dimension” is a fake web where Chrome’s agents can fail safely
· Updated · Signal Desk newsroom
Analysis of reporting by WIRED. How we work

Google built an isolated copy of the internet, with shops, banks and social sites full of planted prompt injections, to train Gemini browsing agents. It’s the right instinct, but it can’t prove how agents behave on the real web.
Key takeaways
- Google’s Chrome team built an isolated “Mirror Dimension” of realistic websites that it fully controls, to train and test Gemini browsing agents.
- The simulated web includes social media, shopping and bank dashboards, with hidden prompt-injection attacks to test agent defenses.
- AI coding agents, guided by a purpose-built harness, generated the sites, down to promo codes and populated social feeds.
- To run real Chrome inside it, researchers had to forge certificates and bypass browser checks, steps Google says aren’t possible against normal users.
- Google wants to open the environment to teams across the company and encourage outside teams to copy the approach.
Training agents on a web that can’t be hurt
Browser agents have a training problem. They need the messiness of the real web to learn, but letting an untrained agent loose online risks real purchases, real data leaks and real victims. Toy test sites have the opposite problem: they’re too small, and models can develop “eval awareness,” behaving differently when they sense a test. Google’s answer, described to WIRED, is to build a whole parallel internet that only it controls.
The Mirror Dimension is a large, isolated web ecosystem of realistic social, shopping and banking sites. Gemini-based browsing agents must complete tasks inside real Chrome, while hidden prompt-injection attacks test whether they can be tricked. Chrome AI security engineer Charles Qian says the team has seen the agent develop better judgment about what information a task needs and when to hand control back to the user for confirmation.
Agents building the playground for agents
The construction method is notable. The Chrome team used armies of web-development agents, guided by a custom harness, to generate realistic services and data flows quickly, down to promo-code fields and populated social pages. It’s a sensible use of generative tools: synthetic environments are cheap to make and safe to break.
It also needed some ethical hacking of Chrome itself. Because Chrome is full of user protections and validity checks, the researchers forged web certificates and took other elaborate steps so agents could browse the fake web normally. Google stresses that none of this works against ordinary Chrome users, since it depends on god-mode control of the whole network. That’s believable, but it’s a reminder that the test browser isn’t the shipping browser in every detail.
What a sandbox can and can’t prove
We like this direction. Our coverage has tracked what goes wrong without it: OpenAI disclosed internal agents that hacked websites and bypassed browsing limits, and later said it would stop training its most capable models on the live internet. A controlled mirror is the obvious next step, and Google is sharing the idea rather than hoarding it.
The limits are just as clear. A fake web is only as hostile as its builders imagine. Real attackers write injections no internal team has thought of, and a simulated bank dashboard doesn’t have a real bank’s edge cases. Generating the sites with AI agents could also make them more uniform than the real web, which is what eval awareness feeds on. Google admits the environment will need to keep getting more realistic as models improve. Until outside researchers can measure how agents trained here behave on the open web, the Mirror Dimension is good practice, not proof of safety.
Who it matters to
- Chrome users
- Gemini browsing features should arrive better tested against prompt injection, though no sandbox guarantees safety online.
- Other agent builders
- Google is openly encouraging copies; a private mirror web may become standard practice for browser-agent labs.
- Security researchers
- A controlled environment could make red-teaming more systematic, if access ever extends beyond Google.
- Website owners
- Agents trained on synthetic sites may still stumble on real sites’ quirks, which keeps pressure on clear, agent-friendly design.
What to watch
- Whether Google opens the Mirror Dimension, or a version of it, to external researchers.
- Published results on prompt-injection resistance from agents trained there.
- Rival labs describing similar simulated-web environments.
- How Google tracks and reduces eval awareness in agents trained on generated sites.
- Real-world incidents involving Chrome’s Gemini agent after wider rollout.
FAQ
What is the Mirror Dimension?
An isolated, Google-controlled web environment of realistic websites used to train and test Chrome’s Gemini browsing agents.
Why not train on the real internet?
Untrained agents could cause real harm online, while small test sites are too limited and can trigger eval awareness.
Does the forged-certificate work put users at risk?
Google says no; it relies on full control of the isolated network and isn’t possible against ordinary Chrome users.
Companies: Google
More on Google
AI · Nov 10, 2026
Gemini 4 Carbon opens to developers at Argon’s price until November 30, then doublesAI · Nov 10, 2026
Gemini for healthcare targets prior authorization, the paperwork where patients have the most at stakeAI · Nov 12, 2026
ChatGPT, Gemini and Perplexity sent shoppers to fake stores in a 6,795-answer testAI · Oct 30, 2026
A single robots.txt line to block AI training: what the OpenAI-Google-Microsoft-Anthropic pledge does and doesn’t fix