CyberEvolver: Self-Evolving Scaffold Architecture for Cybersecurity Agents

arXiv·high signal

Addresses the rigidity of fixed human-designed scaffolds in LLM-based cybersecurity agents by enabling agents to iteratively revise their own scaffold based on execution failures. Tackles three key challenges: unstructured scaffold-change space, sparse execution feedback obscured by environment noise, and error-compounding from low-diversity updates. Practical framework for building adaptive security testing agents.