Dataiku Releases Open-Source Proxy to Mask PII in AI Prompts
Summary
Dataiku’s Kiji Privacy Proxy locally detects and masks PII in requests before forwarding them to external AI APIs, then restores original values in responses. It uses a quantized DistilBERT model with ONNX Runtime.
Key points
- The proxy detects 16 or more PII categories, including email addresses, phone numbers, Social Security numbers, credit card numbers, and IP addresses.
- It substitutes detected PII with dummy values before forwarding requests to external AI APIs, then restores the original values in responses.
- PII detection runs locally using a quantized DistilBERT model through ONNX Runtime, without external calls for the detection step.
- Project documentation reports latency under 100 milliseconds for most requests; the base model scored 94% F1 on an industry benchmark dataset.
- Deployment options include a macOS Electron app, a Linux server binary, and a Chrome extension.
Article Details
- Event Type
- Release of Kiji Privacy Proxy, an open-source local gateway for detecting and masking personally identifiable information in requests to external AI services.
- Impact
- The proxy is designed to reduce exposure of PII by detecting and replacing it locally before requests are forwarded to external AI APIs, then restoring original values in responses. It may help organizations address privacy and compliance obligations; the article reports no security incident or confirmed harm.