Claude Code (anthropics/claude-code-action) runs in GitHub Actions without built-in network restrictions, unlike GitHub Copilot which includes a network firewall by default. This design prioritizes flexibility but shifts security responsibility entirely to the implementation team. The post demonstrates how to add StepSecurity's Harden-Runner to audit all outbound network calls made by Claude Code during CI/CD runs. A real-world example traces Claude Code's full network activity during a Python image generation task, revealing connections to api.anthropic.com, npmjs.org, pypi.org, bun.sh, statsig.anthropic.com, and GitHub APIs — activity invisible in standard CI logs. The key takeaway is that runtime monitoring via Harden-Runner is essential for organizations deploying Claude Code in production pipelines.
Table of contents
Claude Code: Power Without Built-in Network RestrictionsConfiguration Controls: Necessary but Not SufficientImplementing Harden-Runner: Your Security FoundationReal-World Example: Monitoring Claude Code's Network ActivityDetailed Network Activity AnalysisConclusion: The Future of Secure AI DevelopmentSort: