PhishCascade: Confidence-Gated Multi-agent Phishing Detection

Phishing webpages remain one of the most prevalent attack vectors in modern cyber ecosystems, evolving rapidly to evade both rule-based systems and conventional machine learning models. While recent large language model (LLM)–based methods improve semantic reasoning, applying such computationally intensive models uniformly to all inputs creates a significant resource allocation bottleneck, limiting scalability and increasing deployment costs. To address this challenge, we propose \(PhishCascade\) , a hierarchical phishing detection framework built on a confidence-aware cascade. The framework first employs an adaptive URL filtering module that combines deep textual representations with handcrafted statistical features to efficiently filter high-confidence samples. Only samples with elevated uncertainty are forwarded to a selective multi-agent analysis stage. In this stage, specialized LLM-driven agents collaboratively scrutinize the input, leveraging deep semantic understanding to resolve ambiguities that statistical models miss. Experimental results demonstrate that \(PhishCascade\) achieves a strong accuracy-efficiency trade-off, reaching an F1-score of 0.9811 while substantially reducing amortized inference latency compared with monolithic LLM-based pipelines, making it suitable for large-scale, real-time defense.