Online harm detection is crucial for safeguarding digital ecosystems. However, current research is largely confined to single-dimensional classification paradigms or generative safety alignment under simplified and explicit adversarial prompts. To capture complex and highly covert threats in the real world, we introduce the Multi-dimensional Online Harm Detection Benchmark (MDHB). MDHB focuses on four core dimensions critically demanded by social platforms: misinformation, toxicity, sarcasm, and logical fallacies. Notably, we incorporate non-offensive implicit sarcasm (e.g., self-deprecation and veiled complaints) to propel content governance toward proactive psychosocial risk forecasting. Evaluations reveal that mainstream Large Language Models (LLMs), including GPT-5, suffer from severe cross-dimensional interference under both direct inference and ReAct paradigms. To address this, we propose the Self-Refining Sentinel (SRS), an error-driven self-evolving agent architecture. Without requiring parameter fine-tuning, SRS dynamically constructs and updates an experience repository from failure cases through an iterative self-evolution mechanism. Experimental results demonstrate that SRS significantly outperforms state-of-the-art baselines in overall performance, effectively mitigating cross-dimensional interference in multi-dimensional online threat detection.