Thank you for correcting me. I should have said "unmaintained" rather than implied it stopped working. The seed package still runs, you're right. What made me write to worry about is narrower: no real release since mid-2024, and the community fork is currently carrying an open highseverity npm advisory with no one on the hook to patch it. On anonymization, the reason I don't mention it is that Misata is a different category from Snaplet, and your question is the cleanest way to see the line. Snaplet, Tonic, and Greenmask solve "I want a realistic copy of production without the PII". They pull your production data and transform it: mask, subset, format-preserving encryption. The PII risk you're asking about is real for that whole class of tool, because production data is genuinely in the pipeline and the entire job is scrubbing it safely on the way to the dev machine. Misata never reads production at all. It generates data from scratch to fit your schema and constraints. So the direct answer to "how do you prevent production PII from landing on a dev machine" is: production is never in the loop. There is no prod row to leak, because none is ever read. The customer names, emails, and amounts are invented to satisfy the column types and the foreign keys, not derived from anything real. That is a real advantage for some use cases and a real limitation for others, so I'll be straight about the tradeoff. Because it doesn't derive from prod, it will not reproduce your actual distributions, cardinalities, or the weird edge cases that only exist in real data. If your test genuinely needs production shape, a masked copy is the right tool and Misata is not. Where generation wins is when you want data with guaranteed referential integrity and, for testing specifically, a known correct answer: you declare "this month's revenue is exactly X" or "2% of these are fraud" and now you can assert your pipeline computed the right number, which a masked prod copy can't give you because you don't know prod's exact aggregates either. There is an opt-in middle ground (`misata mimic`) that learns the distribution shape from a sample you hand it and emits new synthetic rows rather than copying them, but that's something you deliberately point at a file, not an automatic prod pull, so the clean-hands default stands: no production data unless you choose to involve it.