MedSecLab

MedSecLab is a portfolio-grade reference architecture for securely deploying clinical AI applications in a simulated healthcare environment.

The project connects four related GitHub repositories into one clear story: a security-hardened clinical AI gateway, detection engineering for LLM workloads, adversarial testing, and the homelab infrastructure that ties everything together.

Goal: build a realistic clinical AI security lab using only synthetic healthcare data, then document the architecture, controls, detections, red-team findings, and lessons learned.

Why This Exists

Most homelab projects stop at tool installation: Wazuh, Kasm, OpenEMR, Ollama, or Suricata.

MedSecLab is different. The focus is not simply installing tools. The focus is building a defensible security story around clinical AI:

• How should a local clinical AI system be exposed safely?
• How can synthetic patient records be queried without leaking PHI?
• How can prompt injection and abnormal LLM usage be logged and detected?
• How can red-team findings be mapped to real mitigations?
• How can a small lab simulate enterprise-style healthcare AI security?

Portfolio Repositories

MedSecLab is the umbrella repo. The technical work is split into focused repositories so each hiring audience can quickly evaluate the part they care about.

Repository	Purpose	Status
medseclab	Portfolio landing page, architecture, roadmap, infra notes, lessons learned	Phase 0
clinical-ai-gateway	Secure FastAPI gateway for clinical LLM/RAG workloads	Active
clinical-ai-detections	Wazuh rules, Suricata signatures, dashboards, and detection docs	Active
clinical-ai-redteam	Garak/PyRIT testing methodology, findings, mitigations, and MITRE ATLAS mapping	Planned later

Main Project Tracks

Track 1: Secure Clinical AI Inference Pipeline

Story: I built and secured an end-to-end RAG pipeline that lets clinicians query synthetic patient records using a local LLM, with PHI redaction, audit logging, and OWASP LLM Top 10 controls.

Planned components:

• Data layer: OpenEMR seeded with Synthea synthetic patient data
• Ingestion: Synthea/FHIR records processed through Microsoft Presidio
• Vector database: Qdrant or Chroma
• Inference: Ollama or vLLM serving a local model
• Gateway: FastAPI service with validation, rate limiting, output filtering, and audit logging
• Access layer: Streamlit or React frontend, optionally accessed through Kasm

Primary repository:

clinical-ai-gateway

Main deliverables:

• Working secure AI gateway
• Threat model using STRIDE and optionally LINDDUN
• Security controls mapped to OWASP LLM Top 10, NIST AI RMF, and HIPAA Security Rule technical safeguards
• Demo showing safe querying of synthetic clinical records

Track 2: SOC Detection Engineering for AI Workloads

Story: I developed and tested custom Wazuh rules that detect prompt injection attempts, model exfiltration behavior, and anomalous API usage patterns specific to clinical LLM deployments.

Planned detections:

• Prompt injection signatures
• Role override and jailbreak-style attempts
• Unusual prompt/token volume
• Off-hours access to clinical AI endpoints
• Repeated PHI redaction triggers from one user
• Model file access or tampering anomalies

Primary repository:

clinical-ai-detections

Main deliverables:

• Wazuh decoders and rules
• Example logs and validation tests
• Grafana dashboard: Clinical AI Security Posture
• MITRE ATLAS coverage matrix
• Blog-style detection engineering writeup

Track 3: Adversarial Testing and Hardening

Story: I conducted a structured red-team exercise against my own clinical AI deployment, documented findings using MITRE ATLAS, implemented mitigations, and retested.

Planned testing:

• Garak LLM vulnerability scans
• PyRIT scenarios
• Manual prompt injection tests
• PHI leakage attempts
• Model extraction and abuse-pattern testing against the lab only

Primary repository:

clinical-ai-redteam

Main deliverables:

• Red-team methodology
• Lab-only test scenarios
• Findings report
• MITRE ATLAS mapping
• Mitigations and retest results

Planned Lab Architecture

The final lab simulates a small healthcare provider network. It does not need to run all services at the same time.

Zone	Purpose	Example Services
Clinical	Simulated healthcare user environment	Win11 workstation, OpenEMR, Synthea data
DMZ	Controlled access layer	Kasm, reverse proxy, AI app gateway
SOC	Monitoring and detection	Wazuh, Suricata, Grafana, Loki
AI/ML	Local AI workload	Ollama/vLLM, FastAPI gateway, Presidio, vector DB
Attacker	Red-team testing	Kali Purple, Garak, PyRIT, Atomic Red Team
Mgmt/Infra	Management services	OPNsense, DNS, Vault, Gitea

Repo Layout

medseclab/
├── README.md
├── ARCHITECTURE.md
├── docs/
│   ├── roadmap.md
│   ├── lessons-learned.md
│   ├── compliance-coverage.md
│   └── runbook.md
├── infra/
│   ├── proxmox/
│   ├── ansible/
│   └── networking/
├── diagrams/
│   ├── network.png
│   ├── data-flow.png
│   └── threat-model.png
└── related-repos.md

Important Rules

• No real patient data will be used.
• All healthcare data must be synthetic.
• This is a lab/reference architecture, not a production healthcare system.
• Red-team content targets only the author’s own lab environment.
• Failures and limitations will be documented honestly.

Final Portfolio Outcome

The finished project should include:

• A polished landing repo
• A working secure clinical AI gateway
• Wazuh/SOC detection content for LLM abuse cases
• A structured red-team report
• Architecture diagrams
• Threat model and compliance mapping
• A 5-minute demo video
• Blog posts or writeups explaining the build