claude-plugins-official

Least-privilege permissions

• Before enabling any plugin, review its API scope—verify it requests only the minimum permissions needed for its stated function
• Use API tokens with the narrowest scope available; create dedicated service accounts with read-only access where plugins only need to retrieve data
• Disable unused plugins immediately; don't leave inactive plugins connected with active credentials
• Regularly audit which plugins have access to sensitive systems (databases, billing, authentication services) and revoke access for any non-critical integrations

Network isolation

• Run Claude agents using plugins behind a firewall with egress rules that restrict outbound connections to only approved plugin endpoints
• Use VPN or private endpoints to route plugin API calls; avoid exposing plugin interactions across public networks
• Implement network segmentation so compromised plugins cannot pivot to internal systems or databases outside their intended scope
• Log all outbound connections from Claude agents to detect unexpected plugin API calls to unauthorized destinations

Monitoring

• Set up alerts for unusual plugin API usage patterns—unexpected frequency, data volume, or timing of calls to external services
• Monitor for credential leakage in plugin logs; scan for API keys, tokens, or sensitive data appearing in Claude's conversation history or plugin responses
• Audit plugin responses for command injection or malicious payloads before Claude processes results; validate response format and content against expected schema
• Track plugin version changes and Anthropic security advisories; subscribe to update notifications for any plugins used in production

Red flags to watch for

• Plugins that request broad permissions (e.g., "full account access," "write to any resource") or permissions mismatched to their stated purpose
• Plugins that accept user input directly without explicit sanitization warnings or that lack clear documentation on input validation
• Long gaps between plugin updates or deprecated dependencies; unmaintained plugins are likely vectors for supply-chain attacks
• Plugins that transmit data outside their documented scope (e.g., logging interactions to third-party analytics, sending data to unrelated services)