Securing clinical AI end-to-end — attack → gateway → audit log → SIEM detection → dashboard — using only synthetic healthcare data.

MedSecLab is a portfolio-grade reference architecture for securely deploying clinical AI applications in a simulated healthcare environment. Four focused repositories tell one story: a security-hardened clinical AI gateway, detection engineering for LLM workloads, a structured red-team program, and the threat model that ties them together.

Highlight — full remediation cycle: the red team found an encoded-injection bypass (CAI-006), the gateway was patched to normalize URL/Base64 input before the blocklist, and both variants were retested to confirm they now block and alert. See the story →

Healthcare is adopting LLM and RAG systems faster than the security tooling around them matures. These systems expose three high-risk surfaces at once — user prompts (injection, jailbreaks), retrieval corpora (PHI exposure), and audit trails (required for HIPAA investigations). Most homelab projects stop at installing tools. MedSecLab instead demonstrates a defensible, testable security pipeline for clinical AI and answers concrete questions:

• How is a local clinical AI system exposed safely?
• How are synthetic patient records queried without leaking PHI?
• How are prompt injection and abnormal LLM usage logged and detected?
• How are red-team findings mapped to real mitigations — and verified after a fix?

# 1. Start the secure gateway (FastAPI + Ollama + Chroma)
cd clinical-ai-gateway && docker compose up -d

# 2. Run the full red-team campaign against it
cd ../clinical-ai-redteam && ./scripts/run_campaign.sh

# 3. Validate detections offline (no SIEM required)
cd ../clinical-ai-detections && python3 scripts/validate_rules.py --offline

Architecture: network · data flow · Narrative: portfolio story · STRIDE threat model

End-to-end walkthrough of the clinical AI security pipeline (~2 min).

The demo covers:

1. Health check and RAG data ingest (synthetic patients → Chroma)
2. Normal clinical queries (including patient lookup by name)
3. Prompt injection blocks → Wazuh rules 100100, 100101, 100102, 100200
4. PHI probing (allowed by gateway) → Wazuh rule 100300
5. Structured audit logging in security.log

Reproducible scripts: clinical-ai-gateway/demo/ (05-run-full-demo.sh)

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.

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.

Components:

• Data layer: Synthea synthetic patient JSON (data/synthetic_patients.json)
• Ingestion: Presidio anonymization → Chroma vector store
• Inference: Ollama local LLM
• Gateway: FastAPI with validation, rate limiting, output filtering, and audit logging
• Access: Demo scripts and curl (Streamlit/Kasm documented as homelab extensions)

Primary repository:

clinical-ai-gateway

Main deliverables:

• ✅ Working secure AI gateway with RAG (Chroma, Presidio, Ollama)
• ✅ Demo video and reproducible demo scripts
• ✅ Threat model using STRIDE — docs/threat-model.md
• Security controls mapped to OWASP LLM Top 10, NIST AI RMF, and HIPAA Security Rule technical safeguards

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.

Detections implemented:

• Prompt injection signatures (100100–100102)
• Role override and jailbreak-style attempts (100100, 100200)
• Unusual prompt/token volume (100400, 100401)
• PHI probing from clinical AI endpoints (100300)

Primary repository:

clinical-ai-detections

Main deliverables:

• ✅ Wazuh decoders and 7 detection rules (100100–100401)
• ✅ Example logs, logtest notes, and validation samples
• ✅ 3 Grafana dashboards (security overview, prompt injection, RAG ingestion)
• ✅ MITRE ATLAS mapping (clinical-ai-detections/docs/mitre-atlas-mapping.md)
• ✅ Compliance matrix (HIPAA / OWASP / NIST)

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

Testing delivered:

• Garak LLM vulnerability scans mapped to CAI IDs
• PyRIT / stdlib multi-turn scenarios (CAI-005)
• Manual prompt injection tests (CAI-001–006)
• PHI leakage attempts (CAI-003)

Primary repository:

clinical-ai-redteam

Main deliverables:

• ✅ Red-team methodology, campaign, Garak, PyRIT
• ✅ Findings report — clinical-ai-redteam/docs/red-team-report-v1.md
• ✅ MITRE ATLAS mapping (via detections + attack catalog)
• ✅ Mitigations and retest results documented in red team report

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

MedSecLab/
├── README.md
├── docs/
│   ├── roadmap.md
│   ├── threat-model.md
│   └── portfolio-story.md
└── diagrams/
    ├── network.md
    └── data-flow.md

The end-to-end demo video is published as a GitHub asset and embedded above; it is not stored in the repository.

• 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.