- Home
- Skills
- Legal & Compliance
- GDPR Privacy Scanner: Find Personal Data Before You Build AI
More screenshots
Works with the AI Agents you already use
GDPR Privacy Scanner: Find Personal Data Before You Build AI
Before you train a model, run analytics or share a dataset, you need to know what personal data is in it.
$15
GDPR Privacy Scanner: Find Personal Data Before You Build AI
Example session with this skill installed
We're a small online music shop based in Italy. We want to fine-tune an AI model on our customer and sales database (data/chinook.db) to predict what customers will buy next. Is there personal data in it, what's risky under GDPR, and can you make a safe copy for training? We still need to link customers to their invoices.
- Read your context and instructions
- Compiled the gdpr privacy scanner
Not legal advice — an automated scan plus my review.
What I found: 11 tables, 64 columns. Personal data in 34 columns (scanner: 25; I added 9 on review): 13 direct identifiers (names, addresses, phones, faxes, e-mails), 16 quasi-identifiers, 5 linking IDs. No special categories. 100% of the 59 customers are unique on City + PostalCode.
Biggest risks: contact data in a model that doesn't need it (Art. 5(1)(c), Art. 25); location fields single out customers even without names (Recital 26); purchase histories stay personal data and a model can memorise them (EDPB Opinion 28/2024).
What I did: dropped names, addresses, phones, e-mails, location and staff data; pseudonymised CustomerId with a keyed HMAC so all 412 invoices still join to their 59 customers. Re-scan: direct identifiers 13 → 0, quasi-identifiers 16 → 0, unique records 100% → 0%. Originals unchanged; key in .privacy/secret.key — move it somewhere separate. The output is pseudonymised, not anonymous.
Only you can decide: lawful basis for training (legitimate-interest test), retention, record of processing (Art. 30), a short DPIA, and a processor contract if you fine-tune with an external provider.
Connects securely to your tools. The creator never sees your data.
About this skill
Before you train a model, run analytics or share a dataset, you need to know what personal data is in it. This skill makes your AI coding agent (Claude Code, Codex, Cursor, Gemini CLI) do that check properly, then reduce the data and document it.
What it does
- Scans data: CSV, TSV, Excel, JSON/JSONL, Parquet, SQLite databases and log files. Finds names, e-mails, phones, addresses, IBANs, payment cards and 16 EU and UK national ID numbers, validated by their check digits (codice fiscale, DNI/NIE, NIR, Steuer-IdNr, BSN, PESEL, NIF and more), so order numbers and ISBNs are not mistaken for them.
- Finds the risky data: GDPR Article 9 special categories (health, religion, ethnic origin, trade union, sexual orientation, genetic, biometric), criminal data, children's data, online identifiers and GPS.
- Measures re-identification risk: how many records are unique on ordinary columns like birth date, postcode and gender.
- Reviews code: personal data in logs, sent to LLM APIs or analytics tools, unkeyed hashes of e-mails, special-category fields in schemas, personal data in saved notebook outputs.
- Plans and applies per column: drop, pseudonymise (keyed HMAC-SHA256, so tables still join), generalise (birth year, age band, postcode prefix), mask or redact free text. It tightens the plan until at most 5% of records are unique, applies it on a copy (originals untouched, key kept separately) and re-scans the output.
- Writes a short GDPR report with the article for each finding and masked examples only, plus a plain-language answer: what was found, the biggest risks, what was done, and what only you can decide.
Column names are understood in 22 EU languages.
Tested, with true numbers
- Blind test: a fresh AI agent that never saw the code built 20 tables / 280 columns in 16 languages with labels. The frozen skill found 130 of 164 personal-data columns, with 8 false alarms out of 116 non-personal look-alikes. National IDs 13/13, free text 8/8, e-mails/phones/IBANs/cards/IPs 22/23.
- 9 real public datasets (Chinook, Northwind, Sakila, Titanic, Adult census, heart disease, insurance, Telco churn, wine): answers written before running. Pseudonymising them took direct identifiers from 32 columns to 0 and special categories from 7 to 0.
- Leak check: none of 4,900 real names, e-mails, phones and addresses appeared in any report.
Every result, miss and false alarm is published on the results page.
Not legal advice. It can miss personal data and can flag columns that are not personal; it does not decide your lawful basis. Every report says so. Pseudonymised output is still personal data under the GDPR.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Legal & Compliance.
Reads a sample of up to 20,000 rows per table by default. Cannot read images, audio, PDFs or scans. Misses coded or heavily abbreviated columns (e.g. clinical short names) and is weaker on special-category columns (10/19 in the blind test) and less common languages. Re-identification is measured inside one table. National ID check digits are validated for 14 countries; other countries' IDs are found from column names. It does not decide your lawful basis or replace your DPO.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed today
- Software
Python 3.10+
Permissions
File Scopes
Runs locally. Reads only the files and folders you point it to; writes only to a .privacy/ folder (findings, report, plan, pseudonymised copies, secret key). No network access. Never modifies your original files.