Three principles | Security for Claude Code Development | AI Security Labs

Part 7 of 16

How to control it

Three principles

Least agency: give Claude Code the fewest tools, paths, hosts and credentials that get the job done, and earn autonomy through isolation rather than trust.
Defense in depth: permission rules, the sandbox, hooks, secret hygiene and code review each fail sometimes, so stack them.
Treat all input as untrusted: READMEs, issues, pull request comments, web pages, MCP output and plugin content can all carry instructions aimed at the model.

/* Without JS the drawer cannot open, so show the sidebar inline instead. */ @media (max-width: 900px) { .sidebar { position: static; transform: none; width: auto; box-shadow: none; } .navbar-toggle, .sidebar-head { display: none; } }