Data Leakage Prevention In Skills

Data Leakage Prevention in Skills

Overview

Safeguard against accidental personal data leakage when authoring Hermes skills. Ensures no PII (personally identifiable information), secrets, credentials, or sensitive data is embedded in SKILL.md content, template files, references, or script examples.

When to Use

• Before creating any new skill
• When reviewing/editing existing skills for privacy compliance
• When a skill will process user data
• When writing example templates that might contain sample data

The Risk: How Skills Leak Data

Skills can leak data through multiple vectors:

1. Example content: Templates with hardcoded names, emails, addresses, phone numbers
2. Configuration files: Scripts with embedded API keys, passwords, file paths
3. Reference files: Documents containing sample client data, screenshots with PII
4. Code examples: Python/bash scripts with hardcoded credentials or personal paths
5. Frontmatter: Description or metadata containing project names tied to personal identity
6. Linked files: Templates, scripts, or assets that contain real data instead of placeholders

The Pre-Write Privacy Gate

Always run through this checklist before writing ANY skill content:

Step 1: Identify Data Categories

For every field, variable, or example value, ask:

• Could this identify a real person or entity?

Step 2: Safe Placeholder Standards

Always use placeholders that are unmistakably generic:

# BAD (contains PII-like patterns):
sample_api_key: "sk-live-abc123-john-smith-key"
client_email: "john.doe@company.com"
user_name: "John Smith"

# GOOD (generic placeholders):
api_key: "{{API_KEY}}"
email: "user@example.com"
name: "[FULL_NAME]"

Step 3: Secret Management in Scripts

Never hardcode secrets in skill scripts. Always use:

# BAD:
api_key = "sk-12345abcdef-John-Smith-key"
password = "mypassword123"

# GOOD:
import os
api_key = os.environ.get("API_KEY")
password = os.environ.get("PASSWORD")
# Or: raise ValueError("Set API_KEY environment variable")

Step 4: Generic Example Data Templates

For templates, scripts, and reference files:

• Names: Use "Alice", "Bob", "Charlie" (common test names) or "User One"
• Emails: Use example.com, test.com, or company.example.com
• Addresses: Use "123 Example Street" with "Springfield" as city placeholder
• Phone: "+1-555-010-0000" pattern
• Companies: "ACME Corp", "Example Inc", "Globex Corporation"

Privacy Review Process

Before Skill Publication

1. Scan entire SKILL.md for PII patterns (use pii-detection-and-remediation skill)
2. Scan all linked files (references/, scripts/, templates/, assets/)
3. Check all example values, variable names, default configs
4. Verify no real API keys, tokens, or credentials are included
5. Ensure all placeholder patterns are generic and reusable

For Data-Processing Skills

When a skill will handle actual user data:

1. Document what data the skill collects (explicit ## Privacy Impact section)
2. Document what data the skill stores (none / temporary / persistent)
3. Document what data the skill transmits (external APIs, logging, etc.)
4. Include a ## Data Handling Notice in the skill body

Common Pitfalls

1. Example credentials that look real — "user@company.com" might be a real person's email
2. Hardcoded file paths — "C:\Users\John\Documents\ClientProject" reveals identity
3. Test data from real life — Using your own or clients' real data as examples
4. "Realistic" placeholder patterns — Fake names that happen to match real people
5. Reference files with real data — Screenshots, CSVs with real names/emails
6. Embedding secrets in script examples — "for demo purposes" keys that still work

Verification Checklist

• No full names, emails, phones, addresses in any skill content
• All API keys/passwords replaced with {{ENV_VAR}} or YOUR_KEY_HERE
• File paths use env vars ($PROJECT_DIR) or relative paths
• All example data uses generic placeholders (example.com, Alice/Bob, etc.)
• Reference files scanned for hidden PII
• Scripts use os.environ.get() for secrets, never hardcoded
• Data-handling skills include Privacy Impact and Data Handling Notice sections