Continuous profiling has recently emerged as the ``fourth pillar'' of cloud-native observability, providing detailed stack-trace data that shows function-level behavior in production systems. However, existing profiling tools are mainly used for visualization and manual debugging, with limited support for automated anomaly detection and Root Cause Analysis (RCA). To address this limitation, we introduce ProfRCA, a new framework for automated anomaly detection and RCA using continuous profiling data. We represent call stack samples in profiling as graphs to better capture function relationships and use Graph Neural Networks (GNNs) to detect abnormal patterns. For RCA, we use Large Language Models (LLMs) to analyze anomalous call paths impacted by system calls based on function names and context, providing a clearer diagnosis. Experiments with system call delay faults show that ProfRCA improves RCA accuracy by at least 30\% compared to traditional methods. By using historical data with the Retrieval-Augmented Generation (RAG) technique, ProfRCA strengthens the LLM’s ability to diagnose issues, providing more accurate and useful insights. These results highlight the value of using contextual information in RCA and show how our approach leads to a better understanding of system behavior and anomalies.