HoverBot Tests PII Masking Architecture Before Bot Replies

Handling personally identifiable information before a chatbot responds is a critical challenge for software engineers. For instance, visitors might type messages containing sensitive emails before support bots even ask for an address. During a local test of HoverBot's masking path on November 2, 2026, Presidio altered messages like "Send the return label to shopper@example.com ." into "Send the return label to <EMAIL_1>." prior to the reply step using a synthetic domain.

A practical approach to masking PII requires inspecting messages right at the boundary where the reply model receives them. Relying solely on policies telling models to avoid private details comes too late because raw data has already entered the request payload. Therefore, the development team utilized a self-hosted Presidio analyzer and anonymizer to handle this crucial preliminary step.

Stock photo for illustration only, not from the actual event

When running six separate fictional messages through the local HoverBot masking service using configured categories and thresholds, two target strings—the email and a test credit card number—were successfully replaced. However, a phone-shaped string slipped through unmasked. Meanwhile, three other messages containing names, street addresses, and order numbers remained intact because the reply path was not configured to flag those specific types.

These probes represent a localized test rather than an exhaustive performance rate across real conversations. Automated detection cannot guarantee the discovery of every sensitive data point. The missed phone probe highlights the necessity of reviewing negative cases alongside successful outcomes to ensure support workflows remain secure.