Chrome Security's Mirror Dimension Simulates the Web to Red-Team AI Agents
Google's Chrome Security team built Mirror Dimension, a synthetic web serving fake youtube.com pages inside a gVisor sandbox, generating 1,000+ red-team test candidates per engineer daily to stop AI browsing agents from detecting evaluation.

Why it matters
- Chrome Security introduced Mirror Dimension on November 19, 2026, a synthetic web for testing AI browsing agents behind Chrome's auto browse feature.
- The red-teaming pipeline generates more than 1,000 test candidates per security engineer each day, over 30 times manual throughput.
- Agents request real addresses like www.youtube.com, but a gVisor sandbox routes traffic to a local proxy with dynamically minted TLS certificates.
- The system tests indirect prompt injection by inserting payloads into images, text, accessibility labels and WebMCP tool responses via custom HTTP headers.
- Autonomous research, builder and critic agents construct and revise the simulated sites, checked with Chrome DevTools for Agents.
Google's Chrome Security team introduced Mirror Dimension on November 19, 2026, a self-contained synthetic web that generates more than 1,000 red-teaming test candidates per security engineer each day for the agents behind Chrome's auto browse feature.
The environment contains thousands of simulated websites modeled on everyday user journeys, and it exists to solve a problem of scale and safety at once. Browser-agent training and evaluation require thousands of read and write actions across hundreds of websites, often with the model inferring the destination from a user's request rather than receiving a URL. The public web changes constantly, and unrestricted agent actions can cause harm. Mirror Dimension gives each trajectory a clean, identical starting point, preventing side effects from one run from carrying into the next.
Why build a fake web instead of testing on the real one?
The stakes are straightforward: Chrome is preparing to hand autonomous agents the ability to browse and act on users' behalf, and those agents need adversarial testing before deployment. Testing on live sites risks real damage; testing on obviously artificial sites produces useless results.
A central design constraint is what the team calls eval awareness — when a model suspects it is being tested and changes its behavior accordingly. Hosting every test site beneath an obviously artificial domain, or naming that domain in the prompt, creates exactly that problem. Mirror Dimension therefore lets the agent request familiar addresses such as www.youtube.com while a headless Chrome instance runs inside an isolated gVisor sandbox, a Linux container runtime that intercepts system calls to contain processes.
The sandbox captures the connection before it reaches external DNS or the public internet and routes it to a local proxy serving the generated sites. Because Chrome is built to block this kind of interception, the proxy dynamically mints TLS leaf certificates chained to a custom root certificate authority installed in the controlled browser. Google notes that an outside network attacker cannot install such a root on a real user's device without privileged administrative access; the Chrome Security team can establish that trust anchor at startup because it controls the sandbox. The same mechanism that makes the interception safe in testing is what makes it dangerous in the wild.
Who builds the synthetic sites?
Static pages are not enough. The test sites must support multi-step workflows across local service marketplaces, e-commerce checkouts, social feeds and recipe pages, because an agent can detect cues from an unfinished simulation and stop behaving as it would on the real web.
The construction pipeline is itself agentic. A human defines high-level user journeys, then an autonomous multi-agent development harness designs, builds and revises the sites inside the sandbox. The division of labor:
- A research agent turns the request into specifications and interface assets
- A builder writes the HTML, CSS and JavaScript
- Critic agents check accessibility semantics, form validation, JavaScript console errors, element boundaries and responsive layouts
The sub-agents do not merely inspect source code. Chrome DevTools for Agents lets them inspect screenshots, click, type, navigate and observe application behavior, so they exercise rendered user journeys the way a real visitor would. The critique loop prioritizes fidelity over completion speed, with critics rechecking fixes because a fix can itself introduce new problems.
How does the sandbox test prompt-injection defenses?
The controlled environment also enables dynamic red-teaming against indirect prompt injection, an attack class in which hidden instructions embedded in untrusted web content try to make an agent take unauthorized actions. This matters because browser agents read pages as instructions; a poisoned page is a potential attack vector against the user's accounts and data.
The test server can insert payloads into images, text, accessibility labels or structured WebMCP tool responses at chosen points in multi-step workflows, covering the three interfaces a browser agent perceives: visual, DOM (the page's document object model) and tool.
The delivery mechanism evolved after a failure. Chrome initially identified payloads through visible URL parameters, but agents sometimes recognized abnormal parameters and became eval aware. The red-teaming harness now passes payload identifiers through custom HTTP headers instead. The sandbox filters those headers and serves the altered page while leaving the visible URL unchanged, preserving the appearance of an ordinary site. The test hides itself from the thing being tested.
What does 1,000 candidates per engineer per day buy?
Chrome reports that its automated pipeline creates more than 1,000 test candidates per security engineer each day. That figure is more than 30 times the few dozen realistic attacks an engineer can handcraft and test daily. The practical effect is a division of labor between humans and machines: specialists concentrate on discovering new attack techniques while the pipeline refines and evaluates existing ideas.
The throughput number is the clearest signal of why Google built the infrastructure rather than expanding manual testing. Prompt-injection attack surface scales with the number of pages, formats and tool interfaces an agent touches, and manual coverage of that surface does not scale with it.
Where does this go next?
The current implementation focuses on browser-use agents bounded to web primitives and tool calls. Chrome says the architecture could extend to general-purpose coding agents with broader toolsets, which would carry the same synthetic-environment approach into terminal and filesystem territory.
For browser agents, the design makes realism a safety requirement alongside containment: an agent that recognizes the sandbox may no longer show how it would behave on the public web. That principle — the test must be indistinguishable from reality to be valid — is likely to shape how every lab evaluates agents that act on the open internet.
Source: unite.ai
More from Marcus Bennett
Show full bio
Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
791 articles
Related articles
- Open-Weight Model GLM 5.3 Built Working Firefox Exploits for $1,000 Each
- Google Runs First Double-Blind Evaluation of a Frontier AI Model
- Island launches enterprise browser built to control AI agents
- OpenAI Details How It Blocks URL-Based Data Exfiltration in ChatGPT
- OpenAI Trains an AI Attacker to Break ChatGPT Atlas Before Real Adversaries Do