Skip to content

pii: person names in a name column are not detected (Alice Dupont, Chloé Lambert) #163

Closed

Description

@DavidGodefroid

Problem

A column named name holding person names is not detected as personal data, and is returned in clear by default.

Postgres table customers(id, name, email, phone, iban, vat, city) with rows Alice Dupont, Bob Martin, Chloé Lambert. At the console's first start, the PII scan proposes:

PII: these columns look like personal data and would be masked:
  public.customers: city, email, iban, phone

name is missing, and so is vat (BE VAT numbers, already in lockkit#130). Then SELECT * FROM customers LIMIT 50 returns "Alice Dupont", "Bob Martin", "Chloé Lambert" in clear to the agent.

lockkit's pii is documented to detect "person names (gazetteer)". Either these very common first and last names are outside the gazetteer, or the column scan does not run the name detector on short text columns. The column name itself (name, full_name, nom, naam, customer_name…) is a strong hint that is not used.

Found in the lock family e2e (locksql v0.17.2 image).

Expected

  • The column scan flags name-like columns (column-name context, as in lockkit#130's "context words") when sampled values look like Firstname Lastname.
  • Alice Dupont, Bob Martin, Chloé Lambert (accented names included) are detected by the person-name detector.
  • The column name alone is never enough: a name-like column is flagged only when a significant share of its sampled values is recognised as person names (Firstname Lastname shape, gazetteer first/last names).
  • The table name weighs in: customers, users, employees, contacts, clients strengthen the hint; products, roles, categories, tags, cities weaken it.

Acceptance criteria

  • Bench cases (lockkit#129): name/full_name/nom/naam columns with FR/NL/EN names are flagged.
  • Unit test: accented names (Chloé, Hélène, Zoë) are detected.
  • Negative cases are not flagged: products.name ("Laptop Pro 14"), roles.name ("admin", "editor"), cities.name ("Namur", "Liège"), tags.name ("urgent").
  • locksql's first-start scan proposes public.customers.name on the fixture above.

Found in the local end-to-end test of the lock family (2026-11-09): lox 0.10.0, lockhub v0.10.1, locksql v0.17.2, lockmq/lockfile/lockhttp v0.8.1, lea 0.14.1, fake Postgres/MySQL/RabbitMQ/MinIO/SFTP in Docker, lea agents through MCP.

Activity

  1. DavidGodefroid commented on Nov 9, 2026

    @DavidGodefroid
    OwnerAuthor

    Related: #130 (BE VAT numbers in the same table were not detected either).

  2. DavidGodefroid commented on Nov 9, 2026

    @DavidGodefroid
    OwnerAuthor

    Triage: OK, with a precision: a column called name often has nothing to do with a person. The column name alone must never trigger the detection; the sampled values (and the table name) decide. Negative cases added to the acceptance criteria; the "in doubt, propose the column" line removed.

  3. DavidGodefroid commented on Nov 14, 2026

    @DavidGodefroid
    OwnerAuthor

    lockkit side done: #171 (LooksPersonNameColumn, accented-name and negative unit tests) and #201 (bench name columns: precision 1.000, recall 0.938 over EN/FR/NL/DE, held to the baseline). The last criterion (locksql's first-start scan proposes public.customers.name) is a locksql change, tracked in DavidGodefroid/locksql#236; it needs a lockkit release that includes #201 only for the -er plural table hint.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions