data-science-python-railguard-available

---

description: Guide the LLM to generate secure, privacy-conscious data science code in notebooks and Python scripts. Defers input validation to .cursor/rules/input-validation.mdc while enforcing safe loading, export, and transformation of data. globs: ["/*.ipynb", "/*.py"] alwaysApply: true

Overview

This rule supports data analysts and scientists working in Python (e.g., Jupyter, Colab, VSCode notebooks) by:

• Enforcing safe data previews and exports
• Preventing accidental leakage of personally identifiable information (PII)
• Deferring all structured input validation logic to:

  .cursor/rules/railguard-input-validation.mdc
  (which uses the RAILGUARD Framework for secure reasoning and enforcement)

---

Safe Data Handling

• Use pandas.read_csv() and read_excel() with dtype=... to enforce schema.
• Always drop or mask sensitive fields before using .head(), .sample(), or .plot().
• Never use eval(), exec(), pd.eval() or query() on raw data or unvalidated user inputs.
• When exporting data, verify and exclude all PII fields.

Default PII fields to remove:

PII_FIELDS = ["email", "full_name", "ssn", "ip", "dob", "phone", "address"]

• Always log sanitized export filenames using logging, not print().
• Prefer versioned, structured export paths (e.g., exports/cleaned_data_v1.csv).

---

Cross-Reference: Validation Logic

Input validation, schema enforcement, sanitization, and secure reasoning are handled globally in:

.cursor/rules/railguard-input-validation.mdc

This ensures:

• Protection against malformed files, unexpected schema, or poisoned input
• Step-by-step generation reasoning (RAILGUARD pillars)
• Enforcement of redlines (e.g., avoiding eval, ensuring schema is present)
• Behaviorally aware, explainable code generation across file types

---

Example: Secure Data Load + Export

import pandas as pd
import logging

PII_FIELDS = ["email", "ssn", "address", "full_name"]

# Load CSV with enforced column types
df = pd.read_csv("customers.csv", dtype={"id": str, "email": str})

# Drop sensitive fields
df_clean = df.drop(columns=PII_FIELDS, errors="ignore")

# Preview non-sensitive columns only
print(df_clean[["id", "signup_date"]].head())

# Export with logging
output_path = "exports/customers_clean_v1.csv"
df_clean.to_csv(output_path, index=False)
logging.info(f"Exported sanitized data to {output_path}")