Silent Bug in AI Agent Pipeline Dropped Date Filters Without Any Error

A production AI recruiting assistant built on Azure AI Foundry was silently ignoring date constraints submitted by recruiters, returning incorrect candidate counts without any crash or error signal. The flaw resided in an MCP bridge layer that attempted to extract a typed date parameter from an OData filter string using a regex, which failed whenever the AI model phrased the filter differently. When the regex missed, the date field was sent as null while the raw filter string was still included in the request body, creating a well-formed but semantically broken API call. The application processed the request without complaint, and the agent responded with a confident but factually wrong answer based on an unconstrained dataset. The incident highlights how translation boundaries between AI model outputs and downstream APIs can silently corrupt query intent without triggering any visible failure.