Sandbox

The name comes from the sandbox on the playground. There you're allowed to dig and build, but the sand stays in the box. A sandbox in technology works the same way: a program runs in a separate area of the computer. It gets only limited storage space and is allowed to read only certain files. Everything else on the device remains invisible to the program. If the program crashes or causes damage, only that area is affected. The rest of the system keeps running.

Why this matters

Software you don't fully know is a risk. This applies to downloaded files, to foreign program code, and to email attachments. A sandbox makes it possible to run and observe such things anyway. Security firms do this systematically: they open suspicious files in a sandbox and watch what happens. If the file tries to copy passwords or hide itself, it is classified as malware. On a real machine, this test would have ended disastrously.

With AI, there's a second reason. Modern AI systems don't just write text — they also execute their own program code to solve tasks. This code was generated by a machine, not reviewed by a human. It could contain errors or commands that no one intended. That's why such code almost always ends up in a sandbox.

How it works

Technically, there are several levels. The weakest form is permission management: the operating system denies the program certain kinds of access. Stronger are containers. Here the program gets its own small system environment with its own files, but shares the operating system kernel with the rest. The strictest variant is a virtual machine — a completely simulated computer within the computer. It separates best but requires the most computing power.

In all cases, the same principle applies: there is a boundary, and any access across that boundary must be explicitly permitted. Often network access is also cut off. Then a program cannot send data outward, even if it tries to. A sandbox is never absolutely secure. Time and again, vulnerabilities are found that let attackers break out of the isolated area. Such cases are called sandbox escapes and are considered serious security flaws.

Where you encounter the term

Every browser is a sandbox. Every open webpage runs in isolation so it cannot search through your own hard drive. Apps on your phone work the same way: a flashlight app can't access your photos without asking. In AI products, the sandbox sits behind features like data analysis in ChatGPT or Claude. When a chart is calculated from a spreadsheet there, it runs in an isolated environment on a server.

In financial news, the word has a second meaning as well. Regulatory authorities set up so-called regulatory sandboxes. There, companies are allowed to test new financial services with real customers, under supervision and with relaxed rules. The EU AI Act explicitly provides for such test spaces for AI applications. The principle is the same as in technology: controlled experimentation with limited risk of harm.