CyberEvolver: Self-Evolving AI Agent for Cybersecurity Tasks

CyberEvolver, a newly developed framework for self-evolving cybersecurity agents, has been unveiled by researchers. This system iteratively modifies its structure based on experiences from unsuccessful execution attempts. It tackles the complexities of self-evolution in cybersecurity, where the potential changes to the structure are disorganized, feedback from executions is often limited and obscured, and updates with low diversity can lead to compounded errors. CyberEvolver incorporates a four-layer architecture that breaks down scaffold optimization into manageable components, a trace-to-diagnosis method that transforms noisy logs into useful revision signals, and a population-based beam search approach. This framework is intended for LLM-based agents in cybersecurity, which currently depend on static, human-designed structures that struggle to adapt to various targets and failure scenarios. The research is published on arXiv with the identifier 2605.26195.