Anth Data Handling

Anthropic Data Handling

Overview

Anthropic's data policies: API inputs/outputs are NOT used for model training (commercial API). Zero-day retention is available. This skill covers PII redaction before sending to Claude and compliance patterns.

Anthropic Data Policies

PII Redaction Before API Calls

import re
import anthropic

def redact_pii(text: str) -> tuple[str, dict]:
    """Redact PII before sending to Claude, return redaction map for restoration."""
    redaction_map = {}
    patterns = [
        (r'\b\d{3}-\d{2}-\d{4}\b', 'SSN', '[SSN-REDACTED-{}]'),
        (r'\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b', 'EMAIL', '[EMAIL-REDACTED-{}]'),
        (r'\b\d{3}[-.]?\d{3}[-.]?\d{4}\b', 'PHONE', '[PHONE-REDACTED-{}]'),
        (r'\b\d{4}[- ]?\d{4}[- ]?\d{4}[- ]?\d{4}\b', 'CARD', '[CARD-REDACTED-{}]'),
    ]

    counter = 0
    for pattern, label, replacement in patterns:
        for match in re.finditer(pattern, text):
            counter += 1
            placeholder = replacement.format(counter)
            redaction_map[placeholder] = match.group()
            text = text.replace(match.group(), placeholder, 1)

    return text, redaction_map

def restore_pii(text: str, redaction_map: dict) -> str:
    """Restore redacted PII in Claude's response."""
    for placeholder, original in redaction_map.items():
        text = text.replace(placeholder, original)
    return text

# Usage
user_input = "Contact John at john@example.com or 555-123-4567"
safe_input, redactions = redact_pii(user_input)
# safe_input: "Contact John at [EMAIL-REDACTED-1] or [PHONE-REDACTED-2]"

client = anthropic.Anthropic()
msg = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=256,
    messages=[{"role": "user", "content": safe_input}]
)
final_output = restore_pii(msg.content[0].text, redactions)

Audit Logging

import json
import logging
from datetime import datetime, timezone

audit_logger = logging.getLogger("claude.audit")

def audited_request(client, user_id: str, purpose: str, **kwargs):
    """Wrap Claude API calls with audit logging."""
    # Log request metadata (never log content)
    audit_logger.info(json.dumps({
        "event": "claude.request",
        "timestamp": datetime.now(timezone.utc).isoformat(),
        "user_id": user_id,
        "purpose": purpose,
        "model": kwargs.get("model"),
        "max_tokens": kwargs.get("max_tokens"),
    }))

    response = client.messages.create(**kwargs)

    audit_logger.info(json.dumps({
        "event": "claude.response",
        "request_id": response._request_id,
        "input_tokens": response.usage.input_tokens,
        "output_tokens": response.usage.output_tokens,
        "stop_reason": response.stop_reason,
    }))

    return response

Data Handling Checklist

• PII redacted before sending to Claude API
• Audit logs capture who accessed what and when
• Logs never contain message content or PII
• Data retention policy matches your compliance needs
• Zero-day retention enabled if required (contact Anthropic)
• HIPAA BAA in place if handling PHI
• User consent obtained for AI processing
• Data deletion procedures documented

Error Handling

Prerequisites

• Define the data classification, processing purpose, legal basis or user consent, and retention owner before sending anything to the API.
• Provide an approved redaction policy, a secret-manager-backed API credential, and an allowlisted Anthropic workspace or service boundary.
• Prepare synthetic fixtures that exercise each PII class and a deletion test; do not use real customer records while validating the pipeline.

Instructions

1. Classify the input and reject fields outside the approved purpose or destination. Apply deterministic redaction before constructing the request; keep any restoration map encrypted, access-controlled, and short-lived.
2. Run the redaction, prompt, and response scanners against synthetic fixtures. A failed scan, missing consent, or unexpected content block is a hard stop; do not retry with the original data.
3. Call the Messages API with the least-privileged credential and only the approved model, workspace, and retention configuration. Do not place prompts, responses, redaction maps, or secrets in logs, traces, metrics, or exception text.
4. Scan the response before restoration or release. Record only aggregate counts, policy decisions, request identifier, and token metadata, then enforce the documented retention and deletion procedure.
5. Verify deletion in the sandbox and retain a redacted audit receipt for the owner and compliance reviewer.

Output

Produce a redacted data-handling receipt containing the purpose, policy version, environment, workspace class, redaction and response-scan outcomes, request identifier, token counts, retention deadline, deletion result, and reviewer. Exclude names, contact details, prompt/response text, raw identifiers, redaction maps, and credentials.

Examples

For a synthetic fixture such as customer_id=fixture-017; email=test@example.invalid; purpose=classification, redact the email, call a sandbox workspace, assert raw_pii_sent=0 and sensitive_content_logged=0, and emit redaction=pass; output_scan=pass; retention=24h; deletion=verified. Never substitute a real person or production record in this example.

Resources

• Anthropic Privacy Policy
• Anthropic Security
• Usage Policy

Next Steps

For enterprise access control, see anth-enterprise-rbac.