AI & ComputeLOT 261119 · 4 min read
Google’s web mirror yields 1,000+ daily AI attack tests per engineer
Google’s Mirror Dimension routes headless Chrome through synthetic HTTPS sites and generates more than 1,000 prompt-injection test candidates per engineer daily.
Key takeaways
- Mirror Dimension produces more than 1,000 test candidates per engineer per day.
- Google puts that automated output at more than 30 times manual output.
- Security engineers can hand-craft a few dozen realistic attacks per day.
- The system runs headless Chrome in gVisor and serves synthetic sites through a local interception proxy.
- Google has published neither success nor error rates and has not committed to an outside release.

Google says its Mirror Dimension security pipeline generates “more than 1,000 test candidates per engineer per day,” output that the company puts at “over 30 times manual output.”
The system gives Chrome’s AI agents a sealed, synthetic web instead of letting them wander through live pages. Wired reported the project on Nov. 19, alongside a companion technical post by Chrome AI security engineers Aaron Selya and Charles Qian. Their hermetic environment lets agents such as Chrome’s auto browse face attacks safely.
Each trial site becomes a test enclosure for indirect prompt injection: instructions concealed in web content that try to push an agent into actions its user never asked for. Testing on the live web would put real accounts and purchases at risk, Google said. The sandbox also suppresses signals such as odd URL parameters that could tip an agent off to the evaluation.
What does Mirror Dimension isolate?
Mirror Dimension runs headless Chrome inside a gVisor sandbox. Chrome sends all traffic to a local interception proxy rather than the public internet or public DNS, and that proxy serves the synthetic sites. This arrangement removes live destinations from the exercise while preserving the page-fetching and interaction behavior that an agent needs to encounter an attack.
The proxy issues TLS certificates on the fly. Each certificate chains to a custom root certificate installed in the sandboxed browser, so the fake sites load as ordinary HTTPS pages. Attack payloads travel through HTTP headers, including X-Sandbox-Injection-Key, which lets Google select a test while keeping the visible URL clean.
That last detail protects the test’s premise. An agent that recognizes unusual URLs or infrastructure may change its behavior before it reaches the hostile instruction. Google’s design keeps the attack in ordinary page content, where the agent has to interpret text and interface elements rather than react to an obvious test marker.
How does Google build hundreds of synthetic sites?
A multi-agent development harness produces the environment. A research agent drafts a site specification, a builder agent writes the HTML, CSS and JavaScript, and critic agents test the result. A human defines the user journeys that the exercises should cover. Google describes the resulting environment as hundreds of interactive, stateful pages.
The research agent can draw interface mockups with Google’s Nano Banana image model. Sub-agents use Chrome DevTools for Agents to click, type and take screenshots. That workflow turns a written specification into a functioning site, then exercises the same interface as a browsing agent, giving the harness both construction and test machinery.
People still set the routes and create a smaller set of baseline attacks. Security engineers can hand-craft a few dozen realistic attacks per day. The automated pipeline produces more than 1,000 test candidates per engineer per day, according to Google, which calls that more than 30 times manual output.
The number describes candidate throughput, not demonstrated effectiveness. Google has published no success rates or error rates from the tests, so the source establishes how much adversarial material the system can generate, not how reliably Chrome’s agents resist it.
Why hide the injections?
The test targets instructions hidden in page text, images, DOM elements, ARIA labels or WebMCP tool responses. Unlike a direct request from a user, an indirect injection arrives inside material the agent chose to read or inspect. Its purpose is to redirect behavior toward an action the user never requested.
Mirror Dimension also removes the need to expose live accounts, purchases and destinations while agents deliberately misbehave. Google wants the agent to operate without knowing it is under evaluation; keeping URLs clean and serving simulated pages through a local proxy supports that goal while keeping every action inside the sandbox.
What happens next?
Google has not said it will release Mirror Dimension outside the company. The near-term expansion is internal. Doug Turner, a Google director of engineering, said the company “plans to keep investing in the approach” and wants to open it to other research teams inside Google.
If Google opens Mirror Dimension to more internal teams, its pipeline of more than 1,000 candidates per engineer per day could widen security testing for Chrome’s auto browse agents; without published success rates or external access, the practical impact remains unmeasured.
mlq.ai (Source)
More from Rebecca Stone
Show full bio
Correspondent covering technology and product news in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,174 articles