arXiv:2611.18632v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly being integrated into software security and DevSecOps workflows, yet their adoption, evaluation, and maturity vary substantially across lifecycle stages. This study systematically maps recent LLM-enabled security research published between 2024 and 2026 (until June) across six DevSecOps stages: \textit{Plan}, \textit{Code}, \textit{Test}, \textit{Review}, \textit{Deploy}, and \textit{Monitor}. For each stage, we analyze the security tasks addressed, models and adaptation strategies employed, evaluation settings, reported effectiveness, and key limitations. Our findings reveal an uneven research landscape, with \textit{Test} and \textit{Code} receiving substantially greater attention than \textit{Deploy} and \textit{Review}. Across stages, recent systems increasingly rely on structured context, iterative reasoning, fine-tuning, retrieval augmentation, or agentic workflows rather than direct single-shot prompting. However, reported performance remains difficult to compare because studies use heterogeneous tasks, datasets, metrics, and baselines. Furthermore, most approaches are evaluated using benchmarks or open-source repositories, while industrial-scale evidence remains limited. We also find that existing systems are predominantly stage-specific, with little evidence of LLM-enabled security reasoning spanning multiple DevSecOps stages. Based on these observations, we identify research opportunities in standardized evaluation, production-scale validation, supply-chain and Infrastructure-as-Code security, trustworthy autonomous remediation, and cross-stage LLM integration. This study provides a recent lifecycle-level view of where LLM-enabled DevSecOps research is concentrated, what methodological patterns are emerging, and where significant gaps remain.
Continue to arXiv Software Engineering →You are being forwarded in 6s — stay on this page