security

Sandbox

A fenced-off space where an agent's code can run with limited access, for example to where it can write and which network addresses it can reach. The fence limits the damage if something goes wrong, but only for what actually runs inside it. In Claude Code the sandbox covers shell commands only: file and web tools follow permission rules instead, and MCP servers and hooks run outside it. Like a playpen: it only protects the child who is inside it.

See it in practice: Agent Security

Want to learn more about AI?

Peter Saddington has trained 17,000+ people on agile and AI. Let’s talk.

Work with Peter