A practical security guide for developers using AI coding assistants like Claude Code, Cursor, Copilot, and Windsurf. Covers MCP token scoping (prefer read-only, avoid admin tokens), secrets management (use secret managers like Doppler instead of plaintext .env files), filesystem isolation via Docker or Claude Code's /sandbox command, safe permission allowlisting, vetting MCP extensions for malicious code, defending against prompt injection hidden in READMEs or issue descriptions, and reviewing AI-generated code for backdoors or unexpected network calls. Includes a quick-reference table and statistics on the broader security landscape.
Table of contents
MCP servers and token securitySecrets managementFilesystem isolationAgent execution risks on macOSSafe long-running tasks and permission managementSafely using skills and MCP extensionsPrompt injection attacksValidating AI-generated codeQuick reference cardWhy this matters2 Impressions