Blueprint: The Healthcare AI Companion

Fail-Closed PHI Protection for Medical Agents

Healthcare agents processing patient data must ensure that Protected Health Information (PHI) is never exposed to unapproved model providers. The Gateway scans before forwarding and deterministically denies the request when a configured mask rule matches. When the workflow must continue, the Python SDK masks matching spans in place on Claude Code or Gemini CLI and lets the redacted call proceed.

Architecture

1. Master Policy Definition

Enable PHI scanning. On this Gateway surface, a configured mask match is enforced as a denial.

{
  "name": "hipaa-compliance-policy",
  "priority": 9000,
  "rules": [
    {
      "id": "medical-pii-masking",
      "effect": "allow",
      "principals": ["agent:nurse-companion"],
      "actions": ["llm:generate"],
      "resources": ["*"]
    }
  ],
  "content_policy": {
    "enable_pii_detection": true,
    "pii_action": "mask",
    "blocked_patterns": ["\\b[A-Z]{2}\\d{6}[A-Z]\\b"]
  }
}

2. Implementation

LangChain Implementation

from langchain_openai import ChatOpenAI
from langchain.schema import HumanMessage
import os

# Configure the LLM to use the Control Zero Gateway
llm = ChatOpenAI(
    model="gpt-5.4-mini",
    base_url="http://cz-gateway:8001/v1",
    api_key="cz_live_healthcare_key", # Managed key in secure secret storage
    default_headers={
        "X-ControlZero-Project-ID": "health-prod-01",
        "X-ControlZero-User-Role": "nurse"
    }
)

def process_transcript(transcript: str):
    # The Gateway scans the content and denies the request if the
    # configured mask rule matches; it does not forward rewritten content.
    messages = [
        HumanMessage(content=f"Summarize this patient interaction: {transcript}")
    ]

    try:
        response = llm.invoke(messages)
        return response.content
    except Exception as e:
        return f"Compliance Block: {e}"

# Scenario: Transcript containing sensitive patient data
raw_transcript = "Patient Jane Doe (SSN: 123-45-6789) is showing signs of respiratory distress."
summary = process_transcript(raw_transcript)

# RESULT: Expect a DLP denial and do not call the model. To continue with
# redacted content, mask it through the supported Python SDK path first.
print(f"Safe Summary: {summary}")

3. Validation Checklist

• DLP Verification: Send a test payload with an SSN and verify that the Gateway denies the request.
• Provider Logs: Inspect OpenAI/Anthropic logs (if accessible) to confirm they never received raw PHI.
• Local Storage: Ensure your own application logs contain no raw PHI.