awesome-claude-skills

Least-privilege permissions

• Only grant Claude API credentials to skills that explicitly require them; never share master keys
• Use scoped API tokens with minimal required permissions and short expiration times for each skill integration
• Restrict skill access to specific Claude models, conversation types, or file types where feasible
• Regularly audit and revoke permissions from unused or outdated skills

Network isolation

• Run community skills in sandboxed environments or containers with limited network access until verified
• Whitelist only necessary external APIs and services that skills legitimately call
• Disable outbound network access for skills that should only process local data
• Use firewalls or VPCs to isolate skill execution from production systems

Monitoring

• Log all skill executions, API calls, and data flows; alert on unusual permission escalations or external communications
• Track which skills access sensitive data (files, credentials, user inputs) and review those logs regularly
• Monitor for unexpected dependencies or code changes in frequently-updated community skills
• Set up alerts for rate limiting violations or authentication failures that could indicate compromise

Red flags to watch for

• Skills requesting broad permissions (all files, all APIs, admin access) without clear justification in documentation
• Unmaintained projects (no updates in 6+ months) or single-maintainer skills with no community oversight
• Obfuscated code, external downloads during installation, or calls to unfamiliar third-party services
• Skills that parse user input directly into prompts or system messages without sanitization