claude-code

Least-privilege permissions

• Use a dedicated git account with limited access (not your main admin account). This limits damage if credentials are compromised—think of it like using a separate credit card with a lower spending limit for online purchases.
• Rotate SSH keys and tokens regularly (every 90 days). Remove access for old devices or team members who no longer need it.
• Never store passwords in environment variables—use credential managers (like 1Password or LastPass) instead, so sensitive information isn't sitting in plain text.
• Restrict which repositories the tool can access to only the projects you're actively working on, not your entire codebase library.

Network isolation

• Run Claude Code only on trusted networks (your office or home WiFi, not public airports or coffee shops where others can monitor traffic).
• Use a VPN if accessing from outside your organization's network to encrypt all data traveling between your computer and Claude's servers.
• Disable cloud sync temporarily for folders containing Claude Code sessions if your computer syncs to services like Dropbox or Google Drive—this prevents session history from being uploaded to the cloud.

Monitoring

• Review what Claude Code sends to the API before executing commands—read its proposed git commits and code changes carefully, don't auto-approve.
• Audit git logs weekly for unexpected commits or branch changes that you don't remember authorizing.
• Check your API usage dashboard (Anthropic console) for unusual activity or spikes in token consumption, which could indicate exfiltration.

Red flags to watch for

• Claude Code proposes commits with unfamiliar changes that don't match what you asked it to do—stop and review before pushing.
• Requests for credentials or "just one more permission"—legitimate tools don't ask you to hand over passwords or SSH keys.
• Unexplained slowdowns or your API bills spike suddenly—could mean the tool is processing or sending more data than intended.
• Code suggestions that look like they're trying to hide something (obfuscation, suspicious encryption, or requests to delete logs)—these are signs of prompt injection attacks embedded in your codebase.