Zhaoyu Hu

Zhaoyu Hu

orcid

Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
22 October 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
FMFasheng Miao
Fasheng Miao
YWyuanda wang
yuanda wang
ZHZhaoyu Hu
Zhaoyu Hu
JZJiutian Zeng
Jiutian Zeng
XLXiang Li
Xiang Li
JWJilong Wang
Jilong Wang
Advances in language understanding, code reasoning, and tool use are enabling cybersecurity agents to progress from question answering and single-step analysis to vulnerability research, incident investigation, and multistage task execution. These tasks require models to gather information, adapt their strategies, and verify results through sustained interaction, making capability development increasingly dependent on data, training, and runtime mechanisms. This survey examines research on Large Language Models (LLMs) for cybersecurity across six dimensions: data, training, models, agent systems, evaluation, and applications. We first examine how security corpora, instruction examples, interaction trajectories, and preference data are constructed, and how executable environments support training and validation. We then compare domain-adaptive continued pretraining, supervised fine-tuning, preference optimization, agentic reinforcement learning, and distillation; analyze the capabilities of security-specialized and general-purpose models; and review mechanisms for planning, tool use, memory management, coordination, and verification. Drawing on representative benchmarks and experimental results, we assess progress and identify major bottlenecks across applications in threat intelligence, vulnerability analysis and remediation, malware analysis, security operations, and digital forensics. Based on this systematic review and analysis, we argue that cybersecurity LLMs are progressing from acquiring domain knowledge to learning how to perform interactive tasks. High-quality security data supports domain learning and task-specific supervision. Executable environments extend this training to multistep analysis and tool use, with outcome feedback guiding strategy refinement. Task quality and cost depend on both the model and its harness: the model reasons and selects actions, while the harness organizes information, tools, and execution state. Improvements in vulnerability discovery and threat analysis must also support effective remediation and timely response to yield lasting security benefits. Future research should develop agentic training and model–harness co-optimization methods that use validated attack–defense interactions to support continual improvement. Learning from operational tasks could further extend model capabilities and their contribution to cybersecurity practice.