Google raises a closed Internet to train its AI agents without risking the real world

Google creates a synthetic and isolated Internet to train its AI agents, testing its security against cyberattacks without affecting the real world.

2 minutes

fotonoticia 20261120111350 1920
Add DEMÓCRATA to Google

Published

2 minutes

Google has launched a fully isolated internet to test and perfect the web browsing of its artificial intelligence intelligence agents, preventing these tests from posing a security risk to the real environment.

The technology company has built a synthetic, disconnected replica of the global network, a kind of “sandbox,” with which it can train and evaluate with guarantees the agents that explore web pages. To do so, it has had to solve complex engineering challenges linked to network routing and to the automated creation of sites.

One of the problems that this system addresses is that, if an agent detects that a page may be fake—for example, when it encounters a static site where buttons do not respond—it can infer that it is being subjected to tests. In those cases, it tends to alter its behavior and to carry out actions that it would not take in a natural context.

In addition, it is key to generate in the agent the feeling that it is moving through the genuine web. To do this, Google has designed an interface-less version of Chrome that runs inside an isolated container (gVisor), which completely blocks any connection to the public Internet.

When the agent attempts to access a specific address, such as YouTube (https://www.youtube.com), the system intercepts the network request and redirects it to a local server that hosts the simulated version of that web page. To prevent Chrome from interpreting this redirection as a security failure, the proxy issues valid TLS certificates in real time, tied to a custom root Certificate Authority that has been installed in advance in the sandbox browser. In this way, Chrome accepts the HTTPS connection as if it were legitimate.

The simulated websites are generated through a system made up of three subagents. The first defines the functional requirements (such as persistent shopping carts, filters, or forms) and produces the basic visual elements; the second writes the real code (HTML, CSS, and JavaScript) to set up applications from multiple pages that are fully operational; and the third audits the web to locate JavaScript errors, responsive design issues, or accessibility failures (for example, in ARIA labels).

"The creation of safe and aligned agents requires providing them with a space where they can be supervised while they carry out longer workflows and make decisions in novel scenarios outside of their training," Google has explained in a statement shared on its website.

Thanks to this controlled environment, the company can run thousands of daily security and resilience tests against cyberattacks without exposing real systems or data. This prevents agents from being able to delete information, make purchases by mistake, or be exposed to external attacks.

This testing framework also shields agents against the indirect injection of malicious instructions ('prompts') hidden on a page, which try to alter the agent's original orders in order to induce it to carry out unauthorized actions, as well as against injection that is invisible through HTTP headers.