Adversarial Prompt Disentanglement Framework for LLM Security

A new defense framework called Adversarial Prompt Disentanglement (APD) has been proposed to protect Large Language Models (LLMs) from adversarial prompts that exploit semantic ambiguities. These attacks, including jailbreaking and prompt injection, bypass safety mechanisms and produce harmful outputs. The APD framework proactively identifies and neutralizes malicious components before LLM processing. It integrates three innovations: mutual information-based semantic decomposition to isolate adversarial and benign components, graph-based intent classification using spectral analysis to detect malicious patterns, and a lightweight transformer-based classifier. The framework aims to enhance the integrity and availability of LLMs in security-critical applications.