5.1 Data Classification, PII/PHI Sanitization, and Privacy Controls

5.1 Data Classification, PII/PHI Sanitization, and Privacy Controls

Quick Summary: Deploying Claude in enterprise environments requires robust data governance centered on a 4-tier data classification model (Public, Internal, Confidential, Restricted), systematic PII/PHI sanitization prior to prompt submission, and a clear understanding of data retention and training policies across Claude product tiers. While commercial API and Claude Enterprise accounts do not train on customer data, sensitive identifiers must be masked, tokenized, or scrubbed to satisfy regulatory requirements like GDPR, HIPAA, and SOC 2.

As organizations accelerate the adoption of generative artificial intelligence across core business functions, managing data privacy, regulatory compliance, and information security becomes paramount. Domain 6 of the Claude Certified Associate - Foundations (CCAO-F) syllabus evaluates a practitioner's ability to govern data ingestion, apply mathematical and procedural sanitization techniques, and enforce organizational privacy controls.

---

The Enterprise Shared Responsibility Model

Data security in LLM-powered workflows operates under a Shared Responsibility Model between Anthropic (the AI service provider) and the customer organization:

• Anthropic's Responsibility (Security OF the Model & Infrastructure): Physical and cloud infrastructure security, network perimeter defense, foundational model alignment, encryption in transit (TLS 1.3) and at rest (AES-256), SOC 2 Type II compliance, and strict adherence to contractual data training exclusions for commercial tiers.
• Customer's Responsibility (Security IN the Workflow & Ingestion): Data classification, user access management, pre-prompt PII/PHI sanitization, Data Loss Prevention (DLP) configuration, prompt injection hygiene, and auditing downstream model outputs before operational distribution.

+-----------------------------------------------------------------------------------+
|                    ENTERPRISE SHARED RESPONSIBILITY MODEL                         |
+-----------------------------------------------------------------------------------+
| CUSTOMER   │ • Data Classification (Public / Internal / Confidential / Restricted)|
| CONTROL    │ • Pre-Prompt Sanitization (Masking, Tokenization, Synthetic Data)    |
| (IN Model) │ • Role-Based Access Control (RBAC) & SSO Identity Provisioning       |
|            │ • Human-in-the-Loop Review & Output Quality Governance               |
+────────────┼──────────────────────────────────────────────────────────────────────+
| ANTHROPIC  │ • Model Weights & Base Alignment (Constitutional AI Guardrails)      |
| CONTROL    │ • Infrastructure Encryption (TLS 1.3 In-Transit, AES-256 At-Rest)    |
| (OF Model) │ • Zero-Training Guarantees for Commercial API & Enterprise Tiers     |
|            │ • Operational Abuse Monitoring & Security Compliance (SOC 2, HIPAA)  |
+-----------------------------------------------------------------------------------+

---

The 4-Tier Enterprise Data Classification Framework

To prevent unauthorized disclosure of intellectual property, regulatory non-compliance, or privacy breaches, organizations categorize data into four distinct sensitivity tiers. Each tier defines specific operational rules for Claude interaction.

Exam Rule: On the CCAO-F exam, if a scenario asks whether an analyst may paste raw customer credit card numbers or unmasked patient health records into Claude to summarize a billing dispute, the correct action is never to paste raw Tier 4 data. The analyst must first sanitize the dataset through a local tokenization or redaction pipeline.

---

PII and PHI Identification and Sanitization Techniques

When working with datasets containing sensitive personal data, practitioners must identify and transform protected identifiers before submitting prompts to Claude.

1. Identifying Protected Identifiers

• Personally Identifiable Information (PII): Any information that can be used to distinguish or trace an individual's identity, directly or indirectly (e.g., full name, SSN, driver's license number, home address, personal email, IP address, passport number).
• Protected Health Information (PHI): Under the U.S. Health Insurance Portability and Accountability Act (HIPAA), PHI encompasses any health information linked to any of the 18 Safe Harbor Identifiers (including patient names, geographic data below state level, dates related to an individual, phone numbers, email addresses, medical record numbers, health plan beneficiary numbers, account numbers, certificate/license numbers, vehicle identifiers, device identifiers, URLs, IP addresses, biometric identifiers, full-face photos, and any unique identifying numbers).

2. Core Sanitization Methodologies

Organizations employ four primary sanitization techniques to preserve data utility while eliminating privacy exposure:

[Raw Input]: "Patient Sarah Jenkins (SSN: 123-45-6789, DOB: 1984-03-12) diagnosed with Type 2 Diabetes."

1. MASKING / REDACTION:
   --> "Patient [REDACTED_NAME] (SSN: [REDACTED_SSN], DOB: [REDACTED_DOB]) diagnosed with Type 2 Diabetes."

2. PSEUDONYMIZATION / TOKENIZATION (Reversible via Secure Local Vault):
   --> "Patient PATIENT_UUID_8841 (SSN: SSN_TOKEN_992, DOB: DOB_TOKEN_331) diagnosed with Type 2 Diabetes."

3. SYNTHETIC DATA SUBSTITUTION:
   --> "Patient Marcus Vance (SSN: 987-65-4321, DOB: 1979-11-04) diagnosed with Type 2 Diabetes."

4. AGGREGATION & GENERALIZATION:
   --> "Female patient in age cohort 40-45 diagnosed with Type 2 Diabetes."

---

Data Retention and Training Policies Across Claude Tiers

Understanding how Anthropic handles customer prompts, uploaded knowledge files, and generated outputs across different service tiers is essential for regulatory audit readiness.

Understanding Zero Data Retention (ZDR) and HIPAA BAAs

• Standard 30-Day Operational Abuse Window: By default, Anthropic retains API requests and responses for 30 days in an isolated, encrypted system solely to investigate potential violations of Anthropic's Usage Policy. Access to these logs is restricted to authorized trust and safety personnel responding to automated security alerts.
• Zero Data Retention (ZDR): Highly regulated organizations (such as global investment banks, defense contractors, and healthcare systems) may contractually establish a Zero Data Retention agreement with Anthropic. Under ZDR, Anthropic does not store prompt or completion payloads on disk beyond the transient memory processing window required to stream the response.
• HIPAA Business Associate Agreements (BAAs): Healthcare entities processing PHI through the Anthropic API can execute a formal BAA with Anthropic, certifying that data in transit and at rest meets HIPAA physical, technical, and administrative safeguards.

---

Enterprise Privacy Controls and DLP Integration

To operationalize data governance at scale, enterprise architecture teams embed Claude within a multi-layered security mesh.

1. Pre-Prompt Data Loss Prevention (DLP) Filters

Organizations deploy inline DLP proxies (e.g., Nightfall, Symantec DLP, Palo Alto Prisma) between employee endpoints and Claude interfaces. When an employee attempts to paste content into Claude, the DLP engine scans for regular expression patterns and entropy markers corresponding to SSNs, credit card primary account numbers (PANs), JSON Web Tokens (JWTs), and API private keys. If detected, the proxy blocks the request or automatically replaces the sensitive token with an anonymized placeholder.

2. Claude Project Workspace Hygiene

When configuring Claude Projects for cross-functional teams:

• Audit Project Knowledge Periodically: Designate Project Administrators responsible for quarterly audits of uploaded reference documents to ensure obsolete files or unredacted customer data are purged.
• Enforce Custom Instruction Guardrails: Include persistent privacy directives within the Project Custom Instructions:

<system_privacy_guardrails>
1. Under no circumstances should you generate or request real-world Personally Identifiable Information (PII).
2. If user inputs contain unmasked phone numbers, email addresses, or account numbers, process the analysis using placeholder tokens (e.g., [CLIENT_A], [EMAIL_1]) and remind the user of corporate sanitization policy.
3. Do not include sensitive financial account details in generated Artifacts.
</system_privacy_guardrails>