ProdMask

prodmask

Find out what a plain copy of your production MySQL or MariaDB database would carry into staging, before you make one.

prodmask scan connects read-only and samples a few hundred rows per table. It reports:

• Personal data and secrets:
  • emails, phone numbers, addresses, names, dates of birth, national and tax IDs, device and hardware IDs, push notification tokens;
  • IBANs, card numbers, password hashes, and passwords stored as plain text;
  • API keys and tokens: Stripe, AWS, GitHub, Slack, Discord, Telegram, Google OAuth, Paddle, Postman and about 60 more formats, keys labelled apikey, and passwords inside URLs and Authorization headers (what counts), including inside JSON and PHP-serialized values;
  • email addresses inside URLs, and messages people send each other, such as private messages and notifications.
• What staging would do to real people:
  • payment gateways in live mode;
  • mail settings that reach real inboxes;
  • active webhooks;
  • scheduled subscription renewals and queued jobs.
• How the tables link up: declared foreign keys, plus links inferred from naming and checked against the data, including Laravel polymorphic relations.

It knows WordPress and WooCommerce, Magento 2 and Laravel, and works on any MySQL or MariaDB schema.

Reports name tables, columns, keys and JSON paths, with counts and categories. They never contain the sampled values. Key names can hold data too, such as a user's email address inside a meta key. Those are shown as patterns like um_email_as_username_*, one line for all the keys that share the pattern.

Install

1. Download the archive for your system from the releases page. There are builds for Linux, macOS and Windows, each for amd64 and arm64.
2. Download checksums.txt from the same release.
3. Check the archive against it, changing linux_amd64 to match your archive. On macOS, use shasum -a 256 -c instead of sha256sum -c:

grep _linux_amd64.tar.gz checksums.txt | sha256sum -c

1. Unpack the archive and put prodmask on your PATH.

On Windows, compare the output of Get-FileHash with the line in checksums.txt.

To build it yourself instead, use Go 1.27 or later. The only dependency is github.com/go-sql-driver/mysql. Install it straight from the module:

go install github.com/ProdMask/ProdMask/cmd/prodmask@latest

Or build it from a checkout:

go build -o prodmask ./cmd/prodmask

Use

PRODMASK_PASSWORD='…' prodmask scan --dsn 'mysql://prodmask_ro@db.internal:3306/shop'

The DSN can be a URL (mysql://user:pass@host:port/db?tls=true, or ?socket=/path/mysqld.sock for a socket) or the Go driver form (user:pass@tcp(host:3306)/db). The PRODMASK_DSN environment variable works too. Keep the password in PRODMASK_PASSWORD rather than in the DSN, so it stays out of shell history.

Exit status: 0 done, 1 error, 2 bad usage, 3 found something at or above --fail-on.

In CI, scan the database your staging copy came from, or the copy itself, and fail the build when it still holds personal data:

prodmask scan --dsn "$STAGING_DSN" --fail-on high --format markdown --out prodmask-report.md

What it does to your database

prodmask scan is built to be safe to run against production:

• Session settings:
  • one connection with a read-only session (transaction_read_only) in autocommit mode, so no snapshot or lock outlives a statement;
  • lock_wait_timeout of 5 seconds, so it never queues behind a schema change;
  • a server-side statement timeout (max_statement_time on MariaDB, max_execution_time on MySQL).
• Queries:
  • every sampling query is a SELECT with a LIMIT, ordered by primary key where there is one; key/value lookups (below) have a LIMIT per key instead;
  • text values are cut to 4,096 characters on the server, and settings to 65,536;
  • binary columns are skipped, except binary columns of up to 16 bytes named like IP addresses: those are read as hex in the same sampling query and decoded in memory, and their values are never shown;
  • row counts come from information_schema estimates.
• Key/value tables (wp_options, wp_postmeta, Magento EAV and the like) are read key by key where they can be, so a key on only a few rows is still seen:

  • Small tables: a table of up to 20,000 rows and 32 MB is read once.

  • Larger tables: when its key column is indexed and the server enforces the statement timeout, a larger table has its keys listed with one SELECT DISTINCT. Up to 20 rows per key are then read with indexed lookups, alongside the usual sample of 3,000 rows.

    • Prefix indexes: a table indexed only on a prefix of its key column, as WordPress meta tables are, needs a full scan to list its keys. It's listed only up to 256 MB.

  • Bigger tables, and tables whose keys can't be listed, are sampled, 3,000 rows each, plus lookups for the keys the platform presets know. Without an index on the key column, those lookups are one pass over the table, run only when the table is small or the server enforces the statement timeout.

  • Budgets: a scan looks up at most 3,000 keys and reads at most 32 MB of values per table, and 30,000 keys, 256 MB of values and 1 GB of full-scan listings and passes in all, the main site's tables first.

  • What's left out: keys past these limits, or past the first 200,000 of a table, are judged by name only, and a warning says so. In a sampled table, keys outside the sample that no preset knows aren't seen.

  • Without a preset, a table counts as a key/value table when it has one of these shapes. It's then read the same way: whole, by listing and looking up its keys, or from a 3,000-row window when it's too large for both.

    • key/value, meta_key/meta_value, option_name/option_value, setting_key/setting_value or config_key/config_value;
    • name/value, path/value or param/value, when the table's name has a settings word such as settings or options;
    • any X_key beside an X_value;
    • an X_name beside a text X_value, when X names settings or fields, as in setting_name or field_name, or when the table's name has a settings word.

    An X that names people (member_name, user_key) or data (customer_name, license_key) keeps the table plain. Except for the shapes that need a settings word, the keys must read as identifiers or labels. Keys with a space that don't read as labels may be names. This check runs before any key is looked up:

    • a table keyed by people's names is judged as a plain table instead. Unless it's a translation table, or the column has a finding of its own, its key column gets a low-confidence finding: the starter config suggests hashing it, commented out;
    • the other keys that may be names are shown as (key hidden).
• Small extra checks:
  • a few queries on indexed columns look for pending jobs, active webhooks, scheduled WordPress events, plugin email settings and Contact Form 7 forms;
  • some lookups (IN lists of up to 100 sampled IDs) verify inferred relationships.

Give it a user that can only read:

CREATE USER 'prodmask_ro'@'%' IDENTIFIED BY '…';
GRANT SELECT, SHOW VIEW ON shop.* TO 'prodmask_ro'@'%';

The report warns when the user it runs as could change the database.

What counts as a credential

• A key in an Authorization: Basic credential: the credential is decoded, and its user and password are read for the known key formats. A key sent as either, as Stripe's curl -u sk_live_…: sends it, is reported as that key with its live or test mode, whatever the other part holds. If a real password comes with it, that's reported as well.
• A password with no key in it is critical, unless it reads as an example: a placeholder word such as password, one character repeated, a value that starts with $, %, {, [ or <, or the user's own name, as in admin:admin.
• An empty or missing password: the credential is critical only if its user reads as a random key: a part of 16 or more characters, with two kinds of character and 10 distinct ones, and no spaces, @ or dots. deploybot: doesn't count.
• A password in a URL follows the same rule for examples, read after its percent-escapes, and doesn't count at a host reserved for examples. The host can be a name, a bracketed IPv6 address, a template such as ${DB_HOST}, or empty, as in socket URLs. In a list of hosts, only the first is judged. A user or password can hold an @, written plainly or as %40, so each @ before the path may be the one that ends the password, and it counts if any of those readings does: an address after the URL, such as ,ops@example.com, doesn't make the URL an example.

What doesn't count stays out of the report, though it may be a working credential. The span finders in pkg/detect still find it, with the detail placeholder, for tools that clear credentials rather than count them. A Basic credential that doesn't decode to ASCII text is read only if it holds a key.

Starter config

--config-out prodmask.yml turns the findings into a starting point for masking:

• a strategy per column, meta key or JSON path (fake_email, fake_phone, hash, null, …), with a different one for some rows where a platform needs it, such as the WooCommerce orders among WordPress posts;
• hash instead of null or empty in a column with a unique index, where those would give every row the same value, with a comment that says why;
• one rule per family of keys, written as a pattern such as "um_cache_userdata_*": delete, where * stands for the part that varies. Caches the platform rebuilds get delete.
• a name: rule beside mask: when the key names themselves hold personal data;
• tables to empty;
• the inferred relationships;
• the settings to neutralise.

Weak matches appear commented out, and so do suggestions such as emptying a table that looks like a cache. Keys shown as (key hidden) get no rule, since no rule could name them without printing them; a comment says what they hold. The file's header explains patterns and paths, and docs/config.md describes every section, including the ones prodmask Pro reads. Review every line: the config is a starting point, not a guarantee.

Development

go test ./...

The integration tests build fake WordPress/WooCommerce, Magento 2 and Laravel databases, plus a generic schema no preset knows, and scan them. They run when PRODMASK_TEST_DSN points at a server where the user can create databases:

PRODMASK_TEST_DSN='root:…@tcp(127.0.0.1:3306)/' go test ./...

Their findings are compared with the golden files in pkg/scan/testdata. After a deliberate change, rewrite them with go test ./pkg/scan -update, with PRODMASK_TEST_DSN set, and review the diff line by line.

The scanner engine lives in importable packages under github.com/ProdMask/ProdMask/pkg/.... The command line is pkg/cli, which other builds can add commands to; cmd/prodmask is a thin main on top.

All fixture data is invented:

• phone numbers come from the 555-01xx fiction range;
• IP addresses come from the documentation ranges;
• secret-shaped test strings are assembled at run time.

prodmask Pro

The scanner is free and stays open source. prodmask Pro is a separate paid product, in development. It will:

• make small copies with the real shape and fake customers;
• switch off the settings that let staging email or charge anyone;
• check the result.

Text and Markdown reports end with one line about it. See prodmask.com.

License

Apache License 2.0; see LICENSE and NOTICE.