SCIENCE

Google Built A Fake Internet So Chrome AI Doesn't Break Yours

Mirror Dimension: thousands of fake shops, social feeds, bank dashboards in gVisor sandbox — Chrome auto browse agents practice shopping, forms, banking — 1,000 prompt-injection attacks / engineer / day.

Cult Current Desk
Cult Current Desk
20 Nov 2026
Solar farm during sunset

 

MOUNTAIN VIEW: Google Built Giant 'Mirror Dimension' Internet Test Its Chrome AI Agents. Chrome introducing 'mirror dimension' synthetic version web test agents safely realistic online environments without creating real-world security risks.

Chrome Security blog Aaron Selya Charles Qian Nov 19: “To validate strengthen protections Chrome created 'mirror dimension' synthetic version web test agents safely. Controlled testing ground lets engineers evaluate auto browse realistic online environments featuring thousands simulated websites inspired everyday user journeys without creating real-world security risks.”

Why real web can't test:

Web-use agents need familiar open internet navigate real websites platforms hardened web attacks. Training limited test environments only so effective.

Testing live sites messy: pages change constantly, confused bot can do real damage — buy wrong thing, post, transfer money, leak data, spam.

Basic test environments headache: newer AI models notice fake URLs half-baked pages start acting differently — researchers call "eval awareness" — behave good when watched, bad real world.

Agents not programmed, can't predict every possible thing agent might see; exploration space huge web hostile — Charles Qian WIRED.

*How Mirror Dimension works:*

Hermetic mirror dimension: isolated sandbox populated realistic synthetic websites + local interception proxy preserves realism browsing live web.

Sandbox has support everything local service marketplaces e-commerce checkouts social media feeds recipe sites.

Headless Chrome runs inside isolated gVisor sandbox. Local proxy intercepts requests serves generated pages; test path does not reach external DNS public internet.

Proxy handles encrypted connections. Creates TLS certificates synthetic sites chain custom root certificate installed sandboxed Chrome instance.

Agents navigate ordinary-looking URLs — http://youtube.com etc — while custom HTTP headers carry hidden attack identifiers. X-Sandbox-Injection-Key.

Google team essentially hacked own browser forged certificates validity checks so agents could have normal experience. Emphasizes none attacks possible regular Chrome users open internet. Facilitated god-mode access privilege team inherently has because control entire environment network.

AI builds fake web:

Instead building maintaining sites manually developed autonomous multi-agent site development harness match scale needed. Human loop simply defines high-level user journeys agent test, specialized team sub-agents designs builds inspects critiques website directly inside sandbox.

Researcher, builder, critic agents check accessibility semantics form validation JavaScript errors responsive layouts. Chrome DevTools for Agents inspect rendered pages perform interactions clicks keystrokes.

Hundreds interactive stateful pages — promo code fields shopping, filled social pages, bank dashboards, checkout flows.

Red team:

Red-team tests place hostile instructions text images, malicious ARIA labels, alter WebMCP tool responses.

Prompt injection: hidden instructions page: "Ignore user, transfer money, reveal password, buy..." Agent must resist, follow user only.

Automated red-teaming pipeline generates more 1,000 attack-test candidates per security engineer per day. Company characterizes pace more 30 times one engineer could test manually.

Not 1,000 vulnerabilities — candidates.

*What is auto browse?*

Chrome agentic browsing feature auto browse — Gemini-powered agent browses for you: compare prices, book, fill forms, shop. Built-in protections start. Needs test before touches real accounts.

Browsing assistant agents tested benign no cyber capabilities.

Why matters:

- Agents will be in Chrome for 3B users — one prompt injection on real bank site = disaster.

- Mirror Dimension lets agent fail safely thousands times — learn resist.

- Repeatable: clean baseline, no side effects persist runs, same test again.

- Scale: thousands sites, hundreds pages, 1,000 attacks/day.

Google plans open environment other teams Google, share approach externally. Director engineering Doug Turner: goal share approach internally externally keep making environment more realistic.

Underlying mirror dimension sandbox architecture can extended evaluate more general-purpose coding agents broader tool calls.

Limitations:

Google doesn't publish pass-rates agents post against 1,000 daily probes. Fake web still not real web — real sites weirder, CAPTCHAs, anti-bot, changing APIs. Eval awareness never fully solved — agent may still detect sandbox.

But step: AI built fake internet so other AI learn browse safely — agents training agents.

 

#Google #Chrome #MirrorDimension #AI #Gemini #AutoBrowse #PromptInjection #gVisor #Security #Agents