
Illustration
New tool3 min read
Google tests Chrome's AI agent against attacks in a replica of the web
Google is testing its Chrome agent “auto browse” in a rebuilt version of the web with thousands of simulated sites, where the team plants prompt-injection attacks on purpose, the company said in a blog post on 19 November. It publishes no results; the tool is internal test infrastructure.
Key facts
- On 19 November 2026 Google described a “mirror dimension” in a blog post: a rebuilt web of thousands of sites used to test the agent “auto browse”.
- One security engineer can evaluate over 1,000 test candidates a day, more than 30 times manual work; Google gives no hit rates.
- The tool is described as internal test infrastructure; the post does not mention external availability.
Researched and written with AI, to Luigi's brief, positions and editorial rules. Automatically checked against the sources. How I work →
Google tests its Chrome agent “auto browse” against attacks inside a rebuilt version of the web. In a Google blog post, Aaron Selya and Charles Qian, software engineers on Chrome's AI security team, write that the tests run in headless Chrome inside an isolated gVisor sandbox, whose traffic a local proxy intercepts. The agent visits real addresses such as youtube.com but receives synthetic sites, so the model does not realise it is being tested.
The reason, they write, is that tests on the open web are neither repeatable nor safe: the web keeps changing, and an agent given free rein there can cause real harm. Google describes two technical steps. The proxy mints certificates for the rebuilt sites on the fly, chaining to a custom root certificate authority installed only in the sandbox. The sites themselves are built by a team of AI agents: a research agent writes the specification, a builder agent the HTML, CSS and JavaScript, and critic agents test accessibility, form validation and console errors. A human only defines the user journeys.
Attacks at every layer of the browser
The core is red-teaming. Because Google controls the sandbox web server, the team can plant hidden instructions (prompt injection) in text, images, DOM elements, ARIA labels or WebMCP tool responses, at any step of a multi-step task. At first, URL parameters selected the payload; some models noticed and recognised the test. The identifier now travels in HTTP headers such as “X-Sandbox-Injection-Key”, leaving the address bar clean.
The post gives numbers only for scale: thousands of simulated sites, hundreds of them interactive and stateful. One security engineer can now evaluate more than 1,000 test candidates a day, over 30 times the manual rate of a few dozen realistic attacks a day. Google does not publish results such as how often attacks succeeded, and it is unclear whether the tool will ever be available externally: the post describes it as internal test infrastructure.
Background
In its security architecture post of 8 December 2025, Google calls indirect prompt injection the primary new threat facing all agentic browsers. Nathan Parker of the Chrome security team describes there, among other things, a second checking model (the “User Alignment Critic”), confirmations before purchases and payments, and limits that keep the agent to sites approved for the task (“origin sets”). Google concedes that its prompt-injection classifier cannot flag everything, and pays up to $20,000 for findings in this area. Recently Google gave Gemini access to the Mac, and Apple is weighing, according to Bloomberg, its own Mac agent.
Analysis
The post shows how Google hardens agents, and it exposes a gap: a test in a replica proves only as much as the replica. Google itself says models have noticed test cues, and so invests heavily in sites that give nothing away. How close the synthetic sites come to real shops, portals and sign-in forms cannot be checked from outside, because neither hit rates nor the sites are published.
For companies that want browser agents to handle forms, orders or customer portals, this means a vendor's security work is a precondition, not a free pass. The sensible course is to let agents work at first on accounts with tightly limited rights, keep payments and dispatch under human approval, and rehearse workflows in a test environment with test data before going live, as Google does at scale.
Got AI news? Submit news →
FAQ
What is Google Chrome's “mirror dimension”?
A rebuilt web of thousands of synthetic sites in which Google tests the Chrome agent “auto browse” without touching real sites. Google's blog describes it as internal test infrastructure.
Why doesn't Google test agents on the real web?
Google says the web keeps changing, so tests would not be repeatable, and agents could cause real harm there.
Source
More sources
- Google Chrome Security – Hintergrund: User Alignment Critic, Origin Sets, Bestätigungen, Prämie bis 20.000 US-Dollar.
- Wired – Hinweisgeber über Techmeme (19.11.2026); nicht abrufbar (Zugriff gesperrt), keine Fakten daraus übernommen.



