CyberEvolver: Structured Self-Evolution for Cybersecurity Agents On the Fly

CyberEvolver is a self-evolving cybersecurity agent framework designed to enhance the adaptability of LLM-based agents by iteratively revising their scaffolds based on execution feedback. It employs a four-layer architecture for scaffold optimization, a trace-to-diagnosis mechanism for interpreting execution logs, and a population-based beam search to maintain diverse agent variants. Evaluated on CTF challenges and penetration testing tasks, CyberEvolver demonstrated an average success rate improvement of 13.6% over seed agents and outperformed six human-designed agents, highlighting its potential for advancing adaptive LLM applications in cybersecurity.