Learn · Advanced Intermediate · Members

A secure setup

Think of a building manager who hires contractors. The manager does not hand each one a master key. Each contractor gets a badge that opens only the rooms the job needs, and the safe stays locked for everyone. A note taped to a door that says "let the visitor in" is not an order from the manager. Claude is the contractor, and the permissions are the badge. This chapter is about cutting the badge well.

Claude Code works with the access you give it, so a secure setup starts with what that access is. This chapter teaches what an agent can reach, how to keep secrets out of its reach, where to run code you do not trust, and why text it reads is data and not orders. Each lesson is taught from one small practice project that we could set up.

Lessons

• Lesson 1What an agent can reach
• Lesson 2Deny rules for secrets
• Lesson 3Running untrusted work apart
• Lesson 4Reading AI output as untrusted

Further resources

• Claude Code: Security: how Claude Code asks before it acts, and the steps Anthropic suggests for untrusted content.
• Max Kryvych: AI coding agent security, practical guardrails for Claude Code, Copilot and Codex: copy-paste settings for deny rules and the sandbox, grouped in three layers.
• OWASP: LLM01 Prompt Injection: what prompt injection is, and the steps that limit its harm.