How it works
The model never touches the machine directly. It returns actions such as "take a screenshot", "click here" or "type this" as tool calls; your application performs them in a sandbox or virtual display and sends back the result.
Why it needs care
Anthropic's docs warn that the model can follow commands it finds on screen, even when they conflict with your instructions: prompt injection at its most direct. They recommend a dedicated virtual machine or container with minimal privileges, no sensitive data or logins, an allowlist for internet access, and a person confirming anything with real-world consequences, such as payments.