The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documents tools for executing commands inside isolated microVMs (sandbox::run, sandbox::exec). This is the primary and intended purpose of the skill, providing a secure environment for running untrusted code.
[DATA_EXPOSURE]: The skill provides filesystem operations (sandbox::fs::read, sandbox::fs::write) but these are scoped strictly to the ephemeral overlay filesystem of the isolated sandbox, not the host system.
[NETWORK_OPERATIONS]: The skill describes how to enable networking for sandboxes to connect back to the host engine bus. It specifically warns that this is an opt-in configuration at creation time, following the principle of least privilege.
[REMOTE_CODE_EXECUTION]: While the skill's purpose is to execute code, it uses a virtualization layer (libkrun) to ensure that this code is executed in an isolated, ephemeral environment, protecting the host machine.

iii-sandbox