AI-agent skill · Claude Code · Codex · Cursor · Gemini CLI

Is there personal data in your dataset?

GDPR Privacy Scanner lets your AI coding agent check datasets (CSV, Excel, JSON, Parquet, SQLite, logs) and code for personal data before you build AI or analytics on them. It finds names, e-mails, phones, addresses, IBANs, payment cards and 16 EU and UK national ID numbers (checking their check digits), special-category data such as health or religion, and how many people could be re-identified from ordinary columns like birth date and postcode. Then it plans and applies pseudonymisation on a copy, re-checks the result and writes a short GDPR report.

Not legal advice. The scanner can miss personal data and can flag columns that are not personal. Every report says so and asks you to have your DPO or a privacy professional confirm it. Pseudonymised output is still personal data under the GDPR.
Get it on Agensi · $15 ↗Blind test resultsReal dataGitHub ↗

Blind test: data the scanner had never seen

I froze the skill, then asked a fresh AI agent that had never seen its code to build a realistic test set from scratch: 20 tables, 280 columns in 16 languages (Italian, German, Dutch, Spanish, French, Polish, Portuguese, English, Czech, Romanian, Swedish, Greek, Hungarian, Estonian, Bulgarian, Slovak), as CSV, TSV, JSON, JSONL, Excel, SQLite and a log file. Every column has a correct label, and 116 of the 280 are look-alikes that are not personal data (business contact details, order and booking numbers, product codes, timestamps, aggregates). Version 1.1.0 was scored once, with no changes afterwards.

130 / 164personal-data columns found (79%)
8 / 116non-personal columns wrongly flagged
4,900real names, e-mails, phones and addresses checked: none appeared in any report (all masked)
32 → 0columns of direct identifiers before → after pseudonymising 9 real datasets
Results by kind of data
Kind of dataFound
E-mails, phones, IBANs, cards, IPs22 / 23
National ID numbers13 / 13
Names17 / 22
IDs that link a row to a person9 / 17
Quasi-identifiers (birth date, age, postcode, ...)51 / 62
Special categories (Art. 9) and criminal data10 / 19
Free text containing personal data8 / 8
Every miss and false alarm
TableColumnTruthScanner
abbonati_mobile_2026.csvrif_dispdevice_idnot found
slimme_meter_metingen.jsonlgateway.ipip_addressnot found
reservations_hotels.jsonclient.nomperson_namenot found
gabinet_stomatologiczny.xlsx::Pacjencidata_urdate_of_birthnot found
gabinet_stomatologiczny.xlsx::Pacjenciprocedurahealthnot found
gabinet_stomatologiczny.xlsx::Pacjencilekarzperson_namenot found
gabinet_stomatologiczny.xlsx::Pacjencifield_7healthnot found
farmacia_cartao_cliente.csvn_cartaocustomer_idnot found
farmacia_cartao_cliente.csvmedicamento_receitadohealthnot found
loket_aanvragen.tsvnummerplaatvehicle_platenot found
loket_aanvragen.tsvbehandelaarcustomer_idnot found
prijimaci_rizeni_2026.csvdatum_narozenidate_of_birthnot found
prijimaci_rizeni_2026.csvspecialni_potrebyhealthnot found
candidates_pipeline.csvcurrent_titleoccupationnot found
candidates_pipeline.csvgarda_vettingcriminalnot found
dosare_daune_auto.csvleziunihealthnot found
kyc_privatkunder.csvkundnrcustomer_idnot found
kyc_privatkunder.csvadresspostal_addressnot found
kyc_privatkunder.csvpostnrpostal_codenot found
paraggelies_delivery.csvπελάτηςperson_namenot found
paraggelies_delivery.csvδιεύθυνση_παράδοσηςpostal_addressnot found
paraggelies_delivery.csvΤΚpostal_codenot found
paraggelies_delivery.csvτηλ_καταστήματοςnot personalphone
paraggelies_delivery.csvδιανομέας_idcustomer_idnot found
elearning_beiratkozasok.csvfelhasznalonevcustomer_idnot found
elearning_beiratkozasok.csvkoragenot found
elearning_beiratkozasok.csvmunkamenet_iddevice_idnot found
elearning_beiratkozasok.csvtamogatas_emailnot personalemail
biopank_proovid.tsvnimiperson_namenot found
biopank_proovid.tsvsugugendernot found
biopank_proovid.tsvAPOE_genotüüpgeneticnot found
hr_baltic.db::employeessick_days_ytdhealthnot found
hr_baltic.db::payslipspay_datenot personalincome
anketa_naselenie_2026.csvid_респондентcustomer_idnot found
anketa_naselenie_2026.csvнаселено_мястоcitynot found
anketa_naselenie_2026.csvсексуална_ориентацияsex_life_orientationnot found
faktury_dodavatelov.csviban_dodavatelanot personaliban
faktury_dodavatelov.csvsidlo_ulicanot personalpostal_address
faktury_dodavatelov.csvmestonot personalcity
faktury_dodavatelov.csvpscnot personalpostal_code
faktury_dodavatelov.csvkontakt_emailnot personalemail
faktury_dodavatelov.csvkontaktna_osobaperson_namenot found

History, for honesty: an earlier blind test (18 tables, 256 columns, built the same way) found 94 of 139 with version 1.0.0 and 96 of 139 with 1.0.1, with 15 false alarms out of 117. It was weak on ID columns like KundenNr and on languages it did not know (Greek, Czech, Hungarian). I added 22 EU languages and person-ID rules, then built the new blind test above to measure the new version fairly.

Real public datasets

9 well-known datasets (Chinook (SQLite), Northwind (SQLite), Sakila (SQLite), Titanic (titanic3.csv), UCI Adult census, Cleveland heart disease, US medical insurance, IBM Telco churn, UCI red wine (negative control)): 325 columns. I wrote the correct answer for every column before the first run.

VersionPersonal columns foundFalse alarms
1.0.0 (frozen, first run)78 / 11128 / 214
1.1 (after fixes; not blind)93 / 1115 / 214

What it still misses here: abbreviated medical columns in the heart-disease data (Fbs, Thal, Oldpeak...) and US state columns. After pseudonymising with the default plan, records unique on ordinary attributes in the census data fell from 57% to 1.3%. Small tables (8 employees) stay unique: the report says so and suggests dropping them.

Code review on real projects

It also reads code for privacy problems: personal data in logs, sent to LLM APIs or trackers, plain hashes of e-mails, special-category fields in schemas, personal data saved in notebook outputs. On 3 real open-source projects (miguelgrinberg/microblog, fastapi/full-stack-fastapi-template, mckaywrigley/chatbot-ui) the first version raised 7 findings, of which 1 was real: microblog app/models.py:141 MD5 of the user's e-mail sent to Gravatar (unkeyed hash, PRV006). After fixing the noise it raises 1, the real one.

What your agent does with it

  1. Asks what the data is for (AI training, analytics, sharing) and which columns you need to link records.
  2. Scans data and, if you want, code. Every finding comes with the GDPR article, a confidence level and masked examples only.
  3. Reviews the doubtful findings itself and adds what the scanner cannot see.
  4. Plans actions per column: drop, pseudonymise (keyed HMAC-SHA256), generalise (birth year, age band, postcode prefix), mask or redact free text. It tightens the plan until at most 5% of records are unique.
  5. Applies the plan on a copy, never on your originals, with the key kept separately. Then it re-scans the output.
  6. Answers in plain language: what was found, the biggest risks, what was done (before → after) and what only you can decide.

Good to know

Who it is for

EU companies and teams that want to train AI, run analytics or share data built from customer, patient or employee records.

Runs locally

One Python package, standard library only (Python 3.10+; pandas only for Excel/Parquet). Nothing is uploaded.

Check digits, not guesses

IBANs and ID numbers from 14 countries are validated by their check digits, so order numbers and ISBNs are not mistaken for them.

Honest limits

It reads a sample per table, cannot read images or PDFs, and misses coded or abbreviated columns. It does not decide your lawful basis.

Get it on Agensi · $15 ↗GitHub ↗

Made by Muhammad Shahmeer · Data Science MSc, University of Naples Federico II · © 2026, all rights reserved. Not legal advice. Not affiliated with the EDPB or any data protection authority.