Fault activity and structure are important factors for the assessment of seismic hazards. The Anninghe fault is one of the most active strike‐slip faults in southwestern China but has been experiencing seismic quiescence for M > 4 earthquakes since the 1970s. To understand better the characteristics of its highly locked southern segment, we investigate seismicity and ground motion variability using recently deployed multi‐scale dense arrays. Assisted by machine learning (ML) seismic phase picking and event discrimination models, we first compile a high‐resolution catalog of local seismic events. We find limited earthquakes that occurred on the Anninghe fault, consistent with its generally acknowledged high locking degree. Whereas, most newly detected events appear within off‐fault clusters, among which four are closely related to anthropogenic activities (e.g., mining blasts), and two neighboring faults host the remaining ones. We further apply an ML‐based first‐motion polarity (FMP) classifier and successfully obtain a reliable small earthquake focal mechanism, which agrees well with the geologically inferred north‐south trending and eastward dipping of the Anninghe fault. Analyses of ground motion variations along two across‐fault linear arrays show abrupt changes in FMPs and obvious frequency‐dependent site amplifications near the mapped fault traces. It further suggests that, at finer scales, the damaged Anninghe fault zone may have split into two smaller damaged zones at shallower depths, resulting in a typical “flower‐type” fault structure. The efficient workflow developed in this study can be well applied for the longer‐term monitoring and better characterization of the southern Anninghe fault, or other similar regions.