Google reveals a synthetic web for testing Chrome AI agents without risking real sites
The Mirror Dimension preserves familiar URLs while routing agents to isolated replicas. Google says its attack-testing pipeline produces more than 1,000 test candidates per engineer daily.
Google's Mirror Dimension gives Chrome's auto-browse team a repeatable test bed for workflows and prompt-injection attacks, while keeping agents away from live websites. Chrome requests for ordinary domains are intercepted inside a gVisor sandbox and served generated, stateful sites; development agents build and critique those sites, while hidden HTTP headers select attack payloads without changing visible URLs. Google reports over 30 times as many attack-test candidates as one engineer can produce manually—a throughput measure, not evidence of a corresponding increase in successful defenses. Google plans broader internal adoption, not a public launch.
01
The sandbox starts each evaluation from a clean baseline, avoiding changes to real websites and state carrying over between runs.
02
A custom certificate authority lets Chrome trust the proxy-generated sites; installing that trust anchor on a real user's device would require privileged administrative access.
03
Attack instructions can be embedded in text, images, accessibility labels, and web-tool responses at different points in a task.
Chrome’s AI agents can visit familiar web addresses, fill shopping carts and encounter malicious instructions without reaching the real sites. On November 19, Google disclosed the Mirror Dimension, an isolated synthetic web built to evaluate its auto browse feature. The challenge is not just containing the agents: it is keeping them from realizing they are being tested.
The project grew out of a practical testing problem. Browser agents need to perform thousands of actions across hundreds of websites, including actions that change things. The public web keeps changing, making evaluations hard to repeat, and unrestricted agents can cause harm. Google wanted each run to begin from a clean, identical baseline, with no effects carried over from the previous attempt.
Keeping the address, replacing the destination
Simply putting replica websites at an obvious sandbox address was not enough. Google says that can trigger “eval awareness”: a model recognizes the test setting and behaves differently. Explicitly telling an agent to visit a fake address also breaks the realism of tasks where it must work out which website to use.
Instead, agents navigate to normal addresses such as youtube.com. Chrome runs inside an isolated gVisor sandbox, which intercepts network connections before they reach external address lookups or the public internet. A local proxy—a server that handles the intercepted requests—returns a generated website at the expected address.
That interception would normally trigger Chrome’s security checks. To avoid certificate errors, Google installs a custom trusted certificate authority inside the sandbox. The proxy can then generate certificates that this Chrome instance accepts. Google stresses that installing that trust anchor on a real user’s device would require privileged administrative access; this setup depends on controlling the test environment.
Agents build the sites other agents visit
Realistic addresses solve only part of the problem. Google says static or unfinished pages can also give away a test. Its environment needs interactive sites that preserve state across steps, covering workflows such as shopping checkouts, local service marketplaces, social feeds and recipes.
Editorial illustration for Google reveals a synthetic web for testing Chrome AI agents without risking real sites.
Rather than maintain those sites manually, engineers define high-level user journeys for a team of development agents. A research agent specifies requirements, such as persistent shopping carts and coupon validation. A builder writes the site’s code. Critic agents inspect the result for broken behavior, accessibility problems, layout errors and JavaScript failures.
The agents do more than read source code. Through Chrome DevTools tools, they inspect screenshots, click controls, type into fields and navigate the generated pages. Critics send problems back for fixes and check again, including when a repair creates a new fault.
Making the replica hostile
Controlling the sites also lets engineers plant indirect prompt injections: malicious instructions in untrusted content that try to make an agent take unauthorized actions. Google can insert attacks into text, images, accessibility labels and responses from web tools, at different points during a multi-step task.
Early versions identified attack payloads through visible URL parameters. Google found that models sometimes spotted those unusual additions and recognized the evaluation. The system now passes those identifiers through HTTP request headers, which the local server filters and uses to select the attack. The visible address stays normal.
Google says an engineer can handcraft and test a few dozen realistic attacks daily; verifying and refining them becomes the bottleneck. Its pipeline scales that refinement across sites, leaving engineers to seek new attack techniques. The throughput figure measures test-candidate production, not a demonstrated rate of successful defenses.
The next step is broader internal use, not a public product rollout. Google told WIRED it wants more teams across the company to use the environment. Engineering director Doug Turner said continued investment will be needed to make it more realistic as models become more sophisticated.
Sources
blog.googleMirror Dimension for Agents: Evaluating Browser-Use Agents Without the Web
wired.comGoogle Built a Giant ‘Mirror Dimension’ Internet to Test Its Chrome AI Agents
Reader comments
Newest comments first. Replies stay oldest first.