Chrome Built a Fake Web to Test Auto Browse Against Prompt Injection
A.I. / news
Chrome Built a Fake Web to Test Auto Browse Against Prompt Injection
Two Chrome AI Security engineers describe a gVisor sandbox that serves generated sites at real URLs, with injections chosen by a request header.

Google's Chrome security team has described an internal "synthetic version of the web" that it uses to test its agentic browsing feature, auto browse, against prompt-injection attacks without touching real sites. The post, on the Google Security Blog, is by Aaron Selya and Charles Qian, both Software Engineers on Chrome AI Security.
The page as fetched on November 20 carries no publication date. It announces no public release, open-source plan or outside access, and it reports no benchmark scores.

How the Mirror Dimension is built
Headless Chrome runs inside a gVisor sandbox. Its network traffic goes to a local interception proxy that serves generated sites instead of the public internet, and the proxy mints TLS certificates on demand, chained to a custom root certificate authority installed in the sandboxed browser at startup. The authors say that trick works only because they control the sandbox.
Agents visit real addresses such as youtube.com, not sandbox-specific URLs. The authors say this was to reduce what they call eval awareness, models noticing they are being tested.
| Stage | Who does it | What it produces |
|---|---|---|
| Journey definition | A human | Everyday user tasks to cover |
| Research | Research agent | Site specifications |
| Build | Builder agent | HTML, CSS and JavaScript |
| Review | Critic agents | Checks on accessibility, forms, console errors and responsive layout |
The sub-agents view rendered pages through Chrome DevTools MCP, the Apache-2.0 server maintained by the ChromeDevTools GitHub organisation, which has 58.5k stars. Its README says usage statistics are collected by default and can be switched off with --no-usage-statistics.
The numbers the post gives
The post says evaluations run "thousands of read and write actions across hundreds of websites" continuously, and that the sandbox holds hundreds of interactive, stateful pages.
The red-teaming pipeline produces more than 1,000 attack candidates per day per security engineer. The authors put that at more than 30 times what one engineer could write by hand, where a manual process typically yields a few dozen attacks a day. Both figures are Google's own and unaudited.
Why the sandbox hides its tells
Adversarial payloads for indirect prompt injection can be hidden in visuals, the page DOM, ARIA labels or WebMCP tool responses. They are selected by an HTTP request header, X-Sandbox-Injection-Key, because the authors found agents noticed visible URL parameters.
The concern is documented outside Google. A paper titled WARP, listed at ICML 2026, measured four browser-agent injection benchmarks (DoomArena, WASP, AgentDojo and InjecAgent) across 4,608 trajectories and six unnamed frontier models. It found models read more than half of trajectories on every benchmark as evaluations.
Adding an explicit evaluation cue lowered attack success by 10.9 percentage points on DoomArena, and a length-matched non-evaluation cue did not. The WARP authors, led by Jasmine X Li with Andy Zou among them, built a 50-task replacement on WebArena Verified that reports 80.7% attack success against 22.3% for raw DoomArena and 27.0% for WASP.
- WARP (50 tasks)80.7 %
- WASP27 %
- DoomArena (raw)22.3 %
Source: WARP paper listing, ICML 2026, accessed 2026-11-20
Those figures come from a different team and different attacks, so they do not measure Chrome's setup. They show why a browser vendor would invest in hiding the test.
Limits the authors list
Static or half-finished sites still give cues, and the site-building loop favours fidelity over speed. The certificate trick does not apply to real users. Extending the design to general-purpose coding agents is called a future direction.
Chrome's other on-device AI work is covered in our report on Gemini Nano in Chrome, and agent activity on the desktop in our report on Windows 11 detecting coding agents. The post does not say whether the sandbox or its site generator will be released outside Google.
Sources
More in A.I.
- 01NVIDIA Nemotron 3.5 Bolt Is Called 4GB; Its 4-Bit File Is 6.18 GBThe Gemma 4 E4B derivative wins seven of eight benchmark rows against its base on NVIDIA's own harness, but loses five of eight to Qwen3.5 4B.
- 02Anthropic Prices IPO at $83, and Two Outlets Disagree on the ProceedsFortune reports about $79.4 billion raised on 980 million shares; The Next Web, citing Bloomberg and the FT, reports at least $68 billion on 801.4 million.
- 03Tencent Hy4-flash Card Says Weights Are Out; Launch Coverage Says PlannedThe Hugging Face page lists 220B parameters and a custom licence, while launch coverage gives 224B and a future release, at $0.070 per million input tokens.
- 04Anthropic Says Claude Closed a Black Hole Bound Left Between 8.503 and 12The company's post names one model for the result; the paper's own record names two others, and no outside physicist has reviewed it publicly.