A Hybrid Transformer Automata Framework for Behavioral Anomaly Detection of FraudGPT and Dark LLM Threats

The swift rise of FraudGPT along with other malicious large language models (LLMs) in the Dark Intelligence Continuum has led to a threat landscape that is constantly changing. This chapter presents a hybrid machine learning architecture that combines Transformer-based anomaly detection models with finite state machine (FSM) behavioral analysis to locate and neutralize malicious LLM activities. It also comprises of different evaluation metrics, implementation guidelines and research directions that lead to the advancement of secure LLM ecosystems. Experimental results demonstrate that the hybrid model consistently out performs standalone machine learning and rule-based baselines, achieving high accuracy and robust discriminative performance while maintaining realistic, non-ideal behavior. The FSM component enhances interpretability and enables early detection through explicit modeling of behavioral transitions, addressing key limitations of purely data-driven methods.