claude-skills

Least-privilege permissions

• Enable only the specific skills required for your current task; disable unused skills by default
• Configure skill scopes to restrict suggestions to non-sensitive domains (avoid skills for secrets management, credential generation, or production deployment)
• Use role-based access controls to ensure only authorized developers can invoke high-risk skills (infrastructure, database operations)
• Audit skill permissions quarterly and revoke access when team members change roles

Network isolation

• Run claude-skills in an isolated environment without direct access to production databases, APIs, or deployment systems
• Proxy all external API calls through an allowlist; log and review before execution
• Never pass real credentials, connection strings, or API keys into Claude prompts; use placeholder values or mock environments
• Enforce environment variable separation so development, staging, and production credentials are never accessible to the agent

Monitoring

• Log all skill invocations with input prompts, skill names, and outputs; retain logs for audit trails
• Alert on suspicious patterns: repeated failed skill calls, unusual parameter values, or attempts to invoke disabled skills
• Monitor for credential patterns (API keys, tokens, passwords) in Claude responses and block or flag before logging
• Review skill update changelogs before applying new versions; pin dependency versions and scan for known vulnerabilities

Red flags to watch for

• Prompts that request skills to bypass their documented scope or combine unrelated skills into sensitive workflows
• Responses suggesting unsafe patterns: hardcoding secrets, disabling security checks, or accessing production without approval
• Skill outputs containing credentials, connection strings, or sensitive data that should not be in conversation history
• Skills requesting elevated permissions, network access, or integration with external services not explicitly approved by security