Proactive Shield Against AI-Generated Social Manipulation: The Sentinel Framework

The emergence of sophisticated large language models (LLMs) such as GPT-4 and Llama2 has transformed the way communication occurs, while simultaneously elevating the risks posed by AI-generated social engineering schemes (SEAs). This study introduces ConvoSentinel, an adaptive security framework designed to identify and mitigate SEAs across both standalone messages and broader conversational contexts. Utilizing retrieval-enhanced methodologies, the framework achieves improved precision in detecting harmful intent, surpassing the Llama2-7B two-shot benchmark by 12%. ConvoSentinel incorporates supplementary data and layered analytical techniques to deliver reliable threat identification, providing scalable and efficient real-time defenses. Comprehensive evaluations highlight its flexibility in addressing a variety of attack scenarios, while enabling explanation-driven insights through retrospective analyses. These outcomes establish ConvoSentinel as a crucial innovation in defending against AI-enabled threats. Future studies will investigate hybrid attack-defense dynamics and refine methods for discreet detection to strengthen resilience.