deep-research

Deep-Research MCP Connector Security Hardening Guide

Least-privilege permissions

• Restrict API key access: Grant research tasks only the specific API keys they need—never share master credentials. Rotate keys regularly and revoke unused ones immediately.
• Limit query scope: Configure which data sources, databases, or external services the connector can access; prevent it from querying sensitive internal systems unnecessarily.
• Use read-only credentials: When possible, provide API keys with read-only permissions to prevent the connector from modifying or deleting data.
• Isolate user roles: Allow only authorized team members to trigger deep-research tasks; prevent casual users from initiating expensive or sensitive queries.

Network isolation

• Route through secure channels: Ensure all communication with external LLM APIs uses encrypted connections (HTTPS/TLS) and verified certificates.
• Restrict outbound destinations: Configure firewalls to allow the connector to contact only approved LLM providers and research APIs; block all other external traffic.
• Use a proxy or gateway: Funnel requests through a central security proxy that can inspect and log communications before they reach external services.
• Disable public internet access: Run the MCP server on a private network without direct internet access; use VPN or controlled relay points for outbound requests.

Monitoring

• Log all research queries and results: Maintain complete audit trails (without storing raw credentials) so you can review what was researched and what data returned.
• Alert on unusual patterns: Set up automatic warnings for suspicious activity—such as repeated failed queries, requests for sensitive keywords, or abnormal API usage spikes.
• Track API costs and usage: Monitor spending and request volumes in real-time; unexpected surges may indicate abuse or misconfiguration.
• Review external data sources: Regularly check which websites, APIs, and databases the connector is pulling from; verify they remain trustworthy and updated.

Red flags to watch for

• Credentials appearing in logs or error messages: If API keys, tokens, or passwords show up in system logs or user-facing output, immediately rotate those credentials.
• Unexpected query results or answers: If research responses contain contradictory, malicious, or unusually detailed private information, a prompt injection or data leak may have occurred.
• Requests for "hidden" or "admin" operations: Watch for queries asking the LLM to bypass policies, access restricted data, or perform actions outside its intended purpose.
• Sudden spikes in API consumption or errors: Rapid increases in failed requests, timeouts, or cost may signal abuse, a misconfiguration, or an attacker testing the system's limits.