Breaking and Defending LLM-Powered Social Media Bot Detection Systems

A new arXiv paper (2608.15893v1) examines the security of large language model (LLM)-based social media bot detection systems. The study highlights the ongoing arms race between bots and detection tools, where attackers use adversarial learning and behavior imitation to evade detection. Recent advances in LLMs have improved bot detection through deeper semantic and contextual analysis, but this also introduces new attack surfaces targeting the reasoning and generation mechanisms of LLM-based classifiers. The paper references Anthropic's Claude Code Security as an example of industry tools leveraging LLMs for security-critical decisions, motivating a careful study of their attack surfaces. The research aims to break and defend such systems, focusing on the vulnerabilities introduced by LLM integration. The paper is available on arXiv and was announced as a new submission.