Imported from jeremylongshore/tons-of-skills-marketplace (plugins/saas-packs/openrouter-pack/skills/openrouter-data-privacy/SKILL.md). Install upstream with npx skills add jeremylongshore/tons-of-skills-marketplace --skill openrouter-data-privacy. Copyright stays with the author (MIT).

OpenRouter Data Privacy

Overview

When sending data through OpenRouter to upstream LLM providers, you're responsible for ensuring prompts don't leak PII inappropriately. OpenRouter itself does not train on API data, but each upstream provider has its own data retention and training policies. This skill covers PII detection and redaction, placeholder substitution, provider selection for privacy, and consent tracking.

Prerequisites

• An OpenRouter API key (sk-or-v1-...) exported as OPENROUTER_API_KEY — see the openrouter-install-auth skill for setup
• Python 3.8+ with the OpenAI SDK (pip install openai) — every pattern in this skill is Python
• A sensitivity classification for your workloads (public / standard / sensitive) so privacy_aware_completion() can route each one
• A list of providers your org approves for sensitive data, to plug into provider.order with allow_fallbacks: False

Instructions

1. Start with PII Detection and Redaction: adapt PII_RULES (email, phone, SSN, credit card, sk-or-v1- API keys, IPs) to your data, then run scan_and_redact() on representative inputs and review the findings for false positives.
2. When downstream code needs the original values back, use the Placeholder Substitution Pattern instead of plain redaction — PrivacyProxy.anonymize() before the API call, deanonymize() on the model's reply.
3. Classify each workload and route it via Provider Selection for Privacy: privacy_aware_completion() maps sensitivity to a model plus a provider block (order: ["Anthropic"], allow_fallbacks: False for standard/sensitive).
4. Wire the Privacy Middleware into every call path, choosing block_on_pii=True (raise on detection) or auto_redact=True (scrub and continue) per workload.
5. Apply the Enterprise Considerations: hash logged prompts (SHA-256) for GDPR right-to-erasure, and use BYOK for the most sensitive workloads.

PII Detection and Redaction

import re
from dataclasses import dataclass
from typing import Optional

@dataclass
class PiiScanResult:
    clean_text: str
    findings: list[dict]
    has_pii: bool

PII_RULES = [
    ("email", r'\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b'),
    ("phone", r'\b(?:\+1[-.\s]?)?\(?\d{3}\)?[-.\s]?\d{3}[-.\s]?\d{4}\b'),
    ("ssn", r'\b\d{3}-\d{2}-\d{4}\b'),
    ("credit_card", r'\b(?:\d{4}[- ]?){3}\d{4}\b'),
    ("api_key", r'\bsk-or-v1-[a-zA-Z0-9]+\b'),
    ("ip_address", r'\b(?:\d{1,3}\.){3}\d{1,3}\b'),
]

REPLACEMENTS = {
    "email": "[EMAIL]", "phone": "[PHONE]", "ssn": "[SSN]",
    "credit_card": "[CARD]", "api_key": "[API_KEY]", "ip_address": "[IP]",
}

def scan_and_redact(text: str) -> PiiScanResult:
    """Scan text for PII and return redacted version with findings."""
    findings = []
    clean = text
    for pii_type, pattern in PII_RULES:
        matches = re.findall(pattern, clean)
        for match in matches:
            findings.append({"type": pii_type, "value_prefix": match[:4] + "..."})
        clean = re.sub(pattern, REPLACEMENTS[pii_type], clean)

    return PiiScanResult(clean_text=clean, findings=findings, has_pii=len(findings) > 0)

Placeholder Substitution Pattern

import os, uuid
from openai import OpenAI

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key=os.environ["OPENROUTER_API_KEY"],
    default_headers={"HTTP-Referer": "https://my-app.com", "X-Title": "my-app"},
)

class PrivacyProxy:
    """Replace PII with placeholders before API, restore after."""

    def __init__(self):
        self._map: dict[str, str] = {}

    def anonymize(self, text: str) -> str:
        """Replace PII with unique placeholders."""
        result = scan_and_redact(text)
        if not result.has_pii:
            return text

        # Use deterministic placeholders for consistent replacement
        anonymized = text
        for pii_type, pattern in PII_RULES:
            for match in re.finditer(pattern, anonymized):
                original = match.group()
                if original not in self._map:
                    placeholder = f"[{pii_type.upper()}_{len(self._map)}]"
                    self._map[placeholder] = original
                else:
                    placeholder = next(k for k, v in self._map.items() if v == original)
                anonymized = anonymized.replace(original, placeholder, 1)
        return anonymized

    def deanonymize(self, text: str) -> str:
        """Restore original values from placeholders."""
        result = text
        for placeholder, original in self._map.items():
            result = result.replace(placeholder, original)
        return result

# Usage
proxy = PrivacyProxy()
user_input = "Contact john@example.com or call 555-123-4567"
safe_input = proxy.anonymize(user_input)
# safe_input = "Contact [EMAIL_0] or call [PHONE_1]"

response = client.chat.completions.create(
    model="anthropic/claude-3.5-sonnet",
    messages=[{"role": "user", "content": safe_input}],
    max_tokens=200,
)
# Restore PII in the response if model referenced it
result = proxy.deanonymize(response.choices[0].message.content)

Provider Selection for Privacy

# Force specific provider to control data handling
def privacy_aware_completion(messages, sensitivity="standard"):
    """Route to appropriate provider based on data sensitivity."""

    PRIVACY_CONFIG = {
        "public": {
            "model": "openai/gpt-4o-mini",
            "provider": None,  # Any provider OK
        },
        "standard": {
            "model": "anthropic/claude-3.5-sonnet",
            "provider": {"order": ["Anthropic"], "allow_fallbacks": False},
        },
        "sensitive": {
            "model": "anthropic/claude-3.5-sonnet",
            "provider": {"order": ["Anthropic"], "allow_fallbacks": False},
            # Add PII redaction as mandatory pre-processing
        },
    }

    config = PRIVACY_CONFIG.get(sensitivity, PRIVACY_CONFIG["standard"])
    extra = {}
    if config["provider"]:
        extra["extra_body"] = {"provider": config["provider"]}

    return client.chat.completions.create(
        model=config["model"],
        messages=messages,
        max_tokens=1024,
        **extra,
    )

Privacy Middleware

class PrivacyMiddleware:
    """Enforce privacy policies before every API call."""

    def __init__(self, block_on_pii: bool = False, auto_redact: bool = True):
        self.block_on_pii = block_on_pii
        self.auto_redact = auto_redact

    def process(self, messages: list[dict]) -> list[dict]:
        """Scan and optionally redact PII from all messages."""
        processed = []
        for msg in messages:
            content = msg.get("content", "")
            if isinstance(content, str):
                result = scan_and_redact(content)
                if result.has_pii:
                    if self.block_on_pii:
                        raise ValueError(f"PII detected: {[f['type'] for f in result.findings]}")
                    if self.auto_redact:
                        msg = {**msg, "content": result.clean_text}
            processed.append(msg)
        return processed

Output

The privacy flows in this skill produce:

• A PiiScanResult per scan: clean_text with placeholders substituted, findings (PII type + first-4-chars value prefix per match), and a has_pii flag
• Anonymized prompts like "Contact [EMAIL_0] or call [PHONE_1]" plus the placeholder→original map that deanonymize() uses to restore values in the response
• Chat completions served only by approved providers when the provider.order + allow_fallbacks: False config is applied
• A ValueError listing the detected PII types when PrivacyMiddleware runs with block_on_pii=True

Examples

Scanning a support message before it leaves your infrastructure:

result = scan_and_redact("Contact john@example.com or call 555-123-4567")
print(result.clean_text)  # Contact [EMAIL] or call [PHONE]
print(result.has_pii)     # True
print(result.findings)    # [{'type': 'email', 'value_prefix': 'john...'}, {'type': 'phone', ...}]

To keep the values recoverable, run the same input through PrivacyProxy.anonymize() instead, send the placeholder version to the model, then deanonymize() the reply. More worked examples: references/examples.md.

Error Handling

Enterprise Considerations

• OpenRouter does not train on API data; check each upstream provider's data use policy separately
• Use provider.order + allow_fallbacks: false to ensure data only flows to approved providers
• Implement PII redaction as middleware that runs on every request, not optional per-call
• For GDPR right-to-erasure: don't log raw prompts -- hash them (SHA-256)
• Use BYOK for sensitive workloads so data flows directly to the provider under your account
• Build a data classification system that auto-routes based on sensitivity level

References

• Examples | Errors
• Privacy Policy | Provider Routing