Use Case: Healthcare

The Risk

Clinicians face crushing documentation load. Doctors, nurses, and medical secretaries increasingly turn to AI tools to summarize patient notes, draft discharge letters, or translate medical reports — pasting raw patient records without realizing the data leaves the hospital perimeter.

Typical Shadow AI incidents in healthcare:

Scenario	Data exposed	Regulation triggered
Nurse pastes patient notes into ChatGPT for a summary	Full name, NIR, diagnosis, medications	GDPR Art. 9 (health data), HDS
Doctor uploads a discharge letter to AI for translation	NIR, ICD-10 codes, treating physician	HIPAA §164.514 (US partner hospitals)
Medical secretary asks AI to reformat a prescription	Patient name, DOB, drug names, dosages	GDPR Art. 5, HDS certification obligation
Researcher pastes clinical trial data for analysis	Cohort identifiers, genetic markers	GDPR Art. 9(2)(j), CNIL research guidelines

What Noxys Detects

Tier 1 (browser-side, offline) recognizers for healthcare:

Recognizer ID	Pattern	Example
HLT-001	French NIR (Numéro de Sécurité Sociale)	1 85 12 75 115 422 91
HLT-002	ICD-10 / CIM-10 diagnostic code	J45.9, E11.65
HLT-003	NDA / medical record number (configurable regex)	MR-2024-0047821
HLT-004	Drug name + dosage pattern	Metformine 850mg
PII-001	Full name (NER, Tier 2)	Dr. Sophie Martin
PII-002	Date of birth	12/03/1972

Tier 2 (server-side, Presidio NER) adds:

• Name + NIR + ICD code → confirmed patient record
• Drug name + dosage + patient context → prescription data

Policy Configuration

Recommended policy for a French hospital (HDS-certified environment):

# CISO Console → Policies → New Policy
name: "PHI — Block to Any External AI"
scope:
  groups: ["clinical-staff", "medical-secretaries", "nurses"]
  ou: ["Cardiology", "Oncology", "Emergency"]  # or all OUs

rules:
  - condition:
      risk_signals: ["HLT-001", "HLT-002", "HLT-003", "HLT-004"]
      ai_services: all_external  # block regardless of sanctioned status
    action: Block
    coach_message: >
      Patient health data detected. Sending patient records to any external AI
      service is prohibited under HDS certification and GDPR Article 9.
      Use the anonymization tool first, or contact your DPO.

  - condition:
      risk_signals: ["HLT-001", "HLT-002"]
      ai_services: all
    action: Block  # NIR and ICD codes: absolute block, no exceptions

  - condition:
      risk_signals: ["PII-001", "PII-002"]
      ai_services: unsanctioned
    action: Coach
    coach_message: >
      Patient name or date of birth detected. Ensure this content is
      anonymized before using external AI tools.

Research exception: Add a separate policy scoped to research-staff group with action: Log + redact: true for anonymized cohort queries.

Detection Flow

Doctor opens Claude.ai
  │
  │  Pastes: "Summarise this: Patient Sophie Martin, NIR 1 85 12 75 115 422 91,
  │           admitted 14/05/2026, diagnosis J45.9 (asthma), prescribed
  │           Salbutamol 100mcg..."
  │
  ▼
Noxys Extension (Tier 1, offline, 0.9ms)
  ├─ HLT-001 match: NIR detected
  ├─ HLT-002 match: ICD-10 code detected
  ├─ HLT-004 match: Drug + dosage detected
  └─ Risk score: CRITICAL (3 health signals + NIR)
  │
  ▼
Policy Engine
  └─ Rule: PHI — Block to Any External AI
       Action: BLOCK
  │
  ▼
HTTP request CANCELLED
  │
  ├─ User sees: "Patient health data detected" coaching overlay
  │             with link to anonymization tool
  └─ SIEM / DMP receives: event log
     → No raw prompt stored; NIR never transmitted

Anonymization Workflow (Allow Path)

For legitimate AI-assisted research or drafting, Noxys supports a redact-then-allow flow:

1. Noxys detects PHI in the prompt
2. Policy action: Coach with redaction suggestion
3. User clicks "Anonymize and retry" in the coaching overlay
4. Noxys replaces Sophie Martin → [PATIENT], NIR → [NIR-REDACTED], ICD → preserved (non-identifying)
5. Redacted prompt sent to AI; original event logged

Configure in CISO Console → Policies → Redaction Rules.

Compliance Mapping

Regulation	Requirement	How Noxys satisfies it
GDPR Art. 9	Special category health data — higher protection standard	Absolute block on NIR + ICD codes; no raw health data transmitted
HDS (Hébergeur de Données de Santé)	Health data must stay in certified infrastructure	AI services outside HDS perimeter blocked; approved services logged
HIPAA §164.514	De-identification standard for US partner hospitals	Redaction removes all 18 HIPAA identifiers before AI transmission
CNIL research guidelines	Anonymization required for research AI use	Redact-then-allow flow; anonymization audit trail
GDPR Art. 30	Processing records	Per-event log: user, signals, service, action — no raw data

Evidence for DPO and HDS Auditors

Noxys event records include:

• Risk signal IDs (e.g. HLT-001) — no raw NIR value stored
• Policy version applied
• Whether redaction was offered / accepted / rejected
• Timestamp + pseudonymized user ID

Export path: CISO Console → Reports → GDPR Art. 30 Export or API /v1/events?category=health.