With the widespread deployment of IoT devices, home automation systems now allow users to automate their daily routines. However, extensive interactions pose significant risks to smart homes, particularly because of insecure interaction chains. Recent studies focus on analyzing the sources of anomalies by automation semantics (e.g., open-source automation rules) while overlooking the causal relationships necessary to uncover root causes. To address these issues, we designed a prototype called CausalSafe that automatically model causality-driven interactions to identify suspicious interaction chains causing anomalies in closed-source smart homes. Specifically, CausalSafe constructs an interaction graph by mining causal relationships among devices and leveraging the reasoning capabilities of large language models. This graph enables our method to trace the root causes of user-reported anomalies and alert users to prevent recurrence. We tested CausalSafe on a real-world testbed, and the evaluation shows it performs nearly on par with state-of-the-art methods without requiring strict automation semantics.