Security Architecture

Comprehensive security implementation for enterprise AI applications with PII protection, secrets detection, and compliance logging.

---

Overview

The Empathy Framework implements a defense-in-depth security model with multiple layers of protection:

1. Input Sanitization - PII scrubbing before LLM processing
2. Secrets Detection - Automatic detection of API keys, passwords, tokens
3. Audit Logging - JSONL audit trail for compliance (HIPAA, GDPR, SOC2)
4. Encryption at Rest - AES-256-GCM for sensitive data
5. Access Controls - Role-based access control (RBAC) for wizards

---

Architecture Diagram

┌─────────────────────────────────────────────────────────────┐
│                      User Input                              │
└─────────────────────┬───────────────────────────────────────┘
                      │
                      ▼
┌─────────────────────────────────────────────────────────────┐
│              1. PII Scrubber                                 │
│  ┌─────────────────────────────────────────────────────┐    │
│  │ • SSN, Credit Cards, Phone Numbers                  │    │
│  │ • Healthcare: MRN, Patient ID, DOB, Insurance       │    │
│  │ • Financial: Account Numbers, Routing Numbers       │    │
│  └─────────────────────────────────────────────────────┘    │
└─────────────────────┬───────────────────────────────────────┘
                      │ (Scrubbed Text)
                      ▼
┌─────────────────────────────────────────────────────────────┐
│              2. Secrets Detector                             │
│  ┌─────────────────────────────────────────────────────┐    │
│  │ • API Keys (AWS, Stripe, GitHub, OpenAI)            │    │
│  │ • OAuth Tokens, JWT                                 │    │
│  │ • Private Keys (RSA, SSH)                           │    │
│  │ • Database Connection Strings                       │    │
│  └─────────────────────────────────────────────────────┘    │
└─────────────────────┬───────────────────────────────────────┘
                      │ (Validated Text)
                      ▼
┌─────────────────────────────────────────────────────────────┐
│              3. Audit Logger                                 │
│  ┌─────────────────────────────────────────────────────┐    │
│  │ • User ID, Timestamp, Action                        │    │
│  │ • PII Items Removed, Secrets Detected               │    │
│  │ • JSONL Format for SIEM Integration                 │    │
│  └─────────────────────────────────────────────────────┘    │
└─────────────────────┬───────────────────────────────────────┘
                      │ (Logged)
                      ▼
┌─────────────────────────────────────────────────────────────┐
│              4. LLM Processing                               │
│  ┌─────────────────────────────────────────────────────┐    │
│  │ • OpenAI, Anthropic, Google, etc.                   │    │
│  │ • Receives ONLY scrubbed, validated text            │    │
│  │ • No PII or secrets sent to external APIs           │    │
│  └─────────────────────────────────────────────────────┘    │
└─────────────────────┬───────────────────────────────────────┘
                      │ (Response)
                      ▼
┌─────────────────────────────────────────────────────────────┐
│                   User Response                              │
└─────────────────────────────────────────────────────────────┘

---

PII Scrubbing

Standard PII Patterns

Automatically detected and removed:

Type	Pattern	Example
SSN	\b\d{3}-\d{2}-\d{4}\b	123-45-6789
Credit Card	Luhn algorithm	4111-1111-1111-1111
Phone (US)	\b\d{3}-\d{3}-\d{4}\b	555-123-4567
Email	RFC 5322	user@example.com
IP Address	IPv4/IPv6	192.168.1.1

Healthcare-Specific PHI

For Healthcare Wizards (HIPAA compliance):

Type	Pattern	Example
MRN	\bMRN:?\s*\d{6,10}\b	MRN: 123456
Patient ID	\bPT\d{6,10}\b	PT123456
DOB	\b\d{1,2}/\d{1,2}/\d{4}\b	01/15/1980
Insurance ID	\bINS\d{8,12}\b	INS12345678
Provider NPI	\b\d{10}\b (validated)	1234567890

Implementation Example

from empathy_llm_toolkit import EmpathyLLM
from empathy_llm_toolkit.security import PIIScrubber

# Initialize with security enabled
llm = EmpathyLLM(
    provider="anthropic",
    api_key=os.getenv("ANTHROPIC_API_KEY"),
    enable_security=True  # Enables PII scrubbing
)

# Example with PHI
user_input = """
Patient John Doe (SSN: 123-45-6789, MRN: 987654)
called from 555-123-4567 about diabetes medication.
"""

# Process with automatic PII scrubbing
response = await llm.interact(
    user_id="doctor@hospital.com",
    user_input=user_input,
    context={"classification": "SENSITIVE"}
)

# PHI is automatically removed before sending to LLM
# Audit log records: ['ssn', 'mrn', 'phone', 'name']

---

Secrets Detection

Supported Secret Types

Type	Detection Method	Example Pattern
AWS Access Key	AKIA[0-9A-Z]{16}	AKIAIOSFODNN7EXAMPLE
Stripe API Key	sk_live_[0-9a-zA-Z]{24}	sk_live_...
GitHub Token	ghp_[0-9a-zA-Z]{36}	ghp_...
OpenAI API Key	sk-[0-9a-zA-Z]{48}	sk-...
JWT	Base64 + signature validation	eyJ...
Private Keys	-----BEGIN PRIVATE KEY-----	RSA/SSH keys

Implementation Example

from empathy_llm_toolkit.security import SecretsDetector

detector = SecretsDetector()

code_snippet = """
import openai
openai.api_key = "sk-XXXXXXXXXXXXXXXXXXXXXXXXXXXX"
"""

# Detect secrets
detections = detector.detect(code_snippet)

for secret in detections:
    print(f"⚠️ {secret.secret_type}: Line {secret.line}")
    print(f"   Severity: {secret.severity}")
    print(f"   Recommendation: {secret.remediation}")

# Output:
# ⚠️ OPENAI_API_KEY: Line 2
#    Severity: HIGH
#    Recommendation: Remove from code, use environment variables

---

Audit Logging

Log Format (JSONL)

{
  "timestamp": "2025-11-25T10:30:00Z",
  "event_id": "evt_abc123",
  "user_id": "doctor@hospital.com",
  "action": "llm_interaction",
  "classification": "SENSITIVE",
  "security": {
    "pii_scrubbed": 4,
    "pii_types": ["ssn", "mrn", "phone", "name"],
    "secrets_detected": 0,
    "encryption_used": true
  },
  "performance": {
    "duration_ms": 1234,
    "tokens_used": 500
  },
  "compliance": {
    "hipaa_compliant": true,
    "retention_days": 90
  }
}

Compliance Requirements

Regulation	Retention	Encryption	Audit Trail
HIPAA	90 days minimum	AES-256-GCM required	All PHI access
GDPR	Data subject request	At rest + in transit	All processing
SOC2	180 days	Recommended	All access

Implementation Example

from empathy_llm_toolkit.security import AuditLogger

logger = AuditLogger(
    log_file="/var/log/empathy/audit.jsonl",
    retention_days=90  # HIPAA minimum
)

# Automatically logs all interactions when security is enabled
logger.log_interaction(
    user_id="doctor@hospital.com",
    action="view_patient_record",
    classification="SENSITIVE",
    pii_scrubbed=4,
    secrets_detected=0
)

# Query audit logs
logs = logger.query(
    user_id="doctor@hospital.com",
    start_date="2025-11-01",
    end_date="2025-11-30"
)

print(f"Total interactions: {len(logs)}")
print(f"Total PII scrubbed: {sum(log['security']['pii_scrubbed'] for log in logs)}")

---

Encryption

Data at Rest

AES-256-GCM encryption for sensitive data:

from empathy_llm_toolkit.security import encrypt_sensitive_data

# Encrypt PHI before storing
encrypted_data = encrypt_sensitive_data(
    data={"patient_id": "PT123456", "diagnosis": "Diabetes Type 2"},
    encryption_key=os.getenv("ENCRYPTION_KEY"),  # 32-byte key
    classification="SENSITIVE"
)

# Store encrypted data
database.store(encrypted_data)

# Decrypt when needed (with authorization)
decrypted = decrypt_sensitive_data(
    encrypted_data,
    encryption_key=os.getenv("ENCRYPTION_KEY")
)

Data in Transit

All API communications use TLS 1.2+:

llm = EmpathyLLM(
    provider="anthropic",
    api_key=os.getenv("ANTHROPIC_API_KEY"),
    enable_security=True,
    tls_verify=True  # Enforce TLS certificate validation
)

---

Access Controls

Role-Based Access Control (RBAC)

from empathy_llm_toolkit.wizards import HealthcareWizard
from empathy_llm_toolkit.security import AccessControl

# Define roles
access_control = AccessControl()
access_control.add_role("physician", permissions=["read_phi", "write_phi"])
access_control.add_role("nurse", permissions=["read_phi"])
access_control.add_role("admin", permissions=["read_phi", "write_phi", "view_audit_logs"])

# Check permissions before granting access
if access_control.has_permission(user_role="nurse", permission="read_phi"):
    wizard = HealthcareWizard(llm)
    result = await wizard.process(
        user_input="Patient handoff for bed 312",
        user_id="nurse@hospital.com"
    )

---

Best Practices

✅ Do

1. Always enable security for production: enable_security=True
2. Use environment variables for API keys and encryption keys
3. Review audit logs daily for suspicious activity
4. Implement access controls for sensitive operations
5. Encrypt data at rest for SENSITIVE classification
6. Test PII scrubbing before production deployment
7. Sign BAA agreements with LLM providers (for HIPAA)

❌ Don't

1. Never disable security in production
2. Never commit secrets to version control
3. Never skip encryption for healthcare data
4. Never ignore audit log alerts
5. Never share encryption keys across environments
6. Never bypass access controls for convenience

---

Security Testing

PII Scrubbing Test

import pytest
from empathy_llm_toolkit.security import PIIScrubber

def test_pii_scrubbing():
    scrubber = PIIScrubber()

    text = "Patient SSN 123-45-6789 called from 555-123-4567"
    scrubbed = scrubber.scrub(text)

    # Verify PII removed
    assert "123-45-6789" not in scrubbed
    assert "555-123-4567" not in scrubbed

    # Verify scrubbed items tracked
    items = scrubber.get_scrubbed_items(text)
    assert len(items) == 2
    assert any(item['type'] == 'ssn' for item in items)

Secrets Detection Test

def test_secrets_detection():
    detector = SecretsDetector()

    code = 'api_key = "sk_live_XXXXXXXXXXXXXXXXXXXXXXXXXXXX"'
    detections = detector.detect(code)

    assert len(detections) > 0
    assert detections[0].secret_type == SecretType.STRIPE_KEY

---

See Also

• HIPAA Compliance Guide - Healthcare-specific requirements
• LLM Toolkit API - Security API reference
• Industry Wizards - Domain-specific security
• SBAR Example - Healthcare security in action