nihaixia

Hardening Guide: Nihaixia TCM Agent Skill

Least-privilege permissions

• Restrict output scope: Configure the skill to provide pattern analysis and educational references only—never allow it to generate prescriptions, dosages, or standalone treatment recommendations without explicit user confirmation and licensed practitioner review built into the workflow.
• Limit data access: Prevent the skill from accessing patient records, historical medical data, or personally identifiable health information unless absolutely required; use anonymized case examples for training and reference only.
• Gate integrations carefully: If connecting to pharmacy systems, appointment scheduling, or EHR platforms, require manual handoff steps (a human reviews and approves before data flows) rather than direct system integration.
• Audit API keys and credentials: Rotate authentication tokens monthly; never embed API keys in code or logs; use role-based access so the skill can only call permitted backend services.

Network isolation

• Segregate from production systems: Run the skill in a sandboxed environment separate from patient data systems, payment processing, and critical infrastructure.
• Whitelist external calls only: If the skill needs to fetch reference materials or validate herb databases, explicitly allow only known trusted endpoints; block all other outbound connections by default.
• Encrypt data in transit: Use TLS 1.2+ for all communications; require HTTPS-only for any web-facing endpoints that serve the skill.
• Implement request rate limits: Throttle API calls to prevent abuse (e.g., attackers flooding the system with false symptom patterns to bypass logic or extract training data).

Monitoring

• Log all diagnostic outputs: Record what pattern analysis the skill generates for each query, who triggered it, and timestamp; retain logs for compliance and incident investigation.
• Alert on anomalies: Set up automated warnings if the skill receives unusual input (e.g., injection attempts using meridian or herb names as commands, requests for contraindicated combinations, repeated identical queries from different users).
• Track usage metrics: Monitor query frequency, response times, and which diagnostic formulas are being invoked most often; sudden spikes may indicate misuse or a misconfigured integration.
• Version control training data: Document which edition of teachings, case files, and classical texts are embedded in the skill; flag if outputs diverge from expected patterns (sign of tampering or model drift).

Red flags to watch for

• Users bypassing practitioner review: If customers begin citing the skill's output as standalone medical advice in forums, social media, or clinical notes, immediately escalate and reinforce disclaimers.
• Requests for prescription details: Watch for attempts to extract specific dosages, preparation methods, or treatment timelines; these should always be referred to a licensed TCM provider, not generated by the skill.
• Mixing incompatible frameworks: Alert if users ask the skill to blend nihaixia methodology with unrelated diagnostic systems (Western medicine labs, ChatGPT medical advice); this degrades reliability and creates liability.
• Licensing and IP concerns: Monitor for unauthorized commercial use, republication of copyrighted teachings, or integration into third-party products without proper licensing agreements in place.