Sandboxing
Running AI-controlled code in an isolated environment so it can't affect the host system or production.
Sandboxing is a hard requirement for any agent that runs code. The agent might be wrong; it might be jailbroken; it might just have a bug. The sandbox limits the blast radius.
Common implementations: Docker containers (Claude Code, e2b), browser sandboxes (Anthropic Computer Use), isolated cloud VMs (Modal, Daytona), or process-level isolation. Each has tradeoffs in security, latency, and what the agent can do.
The rule of thumb: never give an agent more access than it needs, and always assume it might do something unintended. Default-deny networking, scoped filesystems, and time-bounded execution are baseline.