The interesting bit isn’t the redaction, it’s the codebook. Tippe’s team turned GDPR Articles 4 and 9 and German trade-secret law into instructions, so the gateway judges the whole context of a request instead of grepping for name-shaped strings. That’s the part I’d steal: a proxy between my tooling and the model that can explain why it pulled something.
What keeps me honest is their finding that privacy and usability don’t always reconcile — strip a dense prompt hard enough and the answer degrades. So I’d run it opt-in on the paths where leakage actually costs me, not globally. A ban is the wrong lever anyway; people use these tools regardless.
The story — Pascal Tippe, a doctoral researcher at Fernuniversität Hagen, and master’s student Michael Maximilian Grötzner built Privacy Gateway, a prototype that detects and strips sensitive content before prompts reach a language model. It evaluates full request context rather than keywords, guided by a codebook derived from GDPR Articles 4 and 9 and German trade-secret law. Testing across everyday scenarios showed some prompts can’t be anonymized without degrading answer quality. (Source)