On Predicting Vulnerability Severity Using In-Context Learning: An Industrial Case Study

Until April 2026, the National Vulnerability Database labeled nearly every published Common Vulnerabilities and Exposures (CVE) entry with a Common Vulnerability Scoring System (CVSS) v3.1 vector, a severity tier, and a Common Weakness Enumeration (CWE) weakness type. Now, NIST enriches only the high priority CVEs, leaving many newly published vulnerabilities unlabelled with the classifiers that security teams use for patch triage. In CVE enrichment, we predict the eight base components of CVSS v3.1 vector, base score, severity tier, and CWE from the vulnerability's textual description alone. To prevent leakage between near-duplicate descriptions, all systems are trained only on CVEs published in 2022 and 2023. Results are evaluated on all 36,325 usable 2024 CVEs in the pinned NVD snapshot. All scores are either directly predicted, or calculated from predicted vectors using the CVSS formula. A fine-tuned 3.8B-parameter model Phi-3.5-mini, trained with LoRA in 2.5 hours on a single Apple M5 Pro chip, achieves 69.9% accuracy on severity and 70.9% on CWE. It matches the official vector exactly for 48.8% of test CVEs, outperforming TF-IDF logistic regression by 2.8 to 4.7 points on all accuracy metrics based on zero-exclusion paired bootstrap intervals. When we evaluate the base model with a zero-shot prompt, it achieves 20.8% accuracy on severity in a sample of 2,000 CVEs. In this same sample, its stated scores match its own vector precisely for only 3.4% of CVEs. However, deriving scores from predicted vectors boosts the critical recall of TF-IDF logistic regression from 50.7% to 73.3% with no additional cost. And since the errors of the fine-tuned model and TF-IDF affect different CVEs, at least one system matches the correct severity tier for 75.8% of test CVEs.

Automated vulnerability detection is a fundamental task in software security, yet existing learning-based methods still struggle to capture the structural dependencies, domain-specific vulnerability knowledge, and complex program semantics required for accurate detection. Recent Large Language Models (LLMs) have shown strong code understanding ability, but directly prompting them with raw source code often leads to missed vulnerabilities or false alarms, especially when vulnerable and benign functions differ only in subtle semantic details. To address this, we propose VulTriage, a triple-path context augmentation framework for LLM-based vulnerability detection. VulTriage enhances the LLM input through three complementary paths: a Control Path that extracts and verbalizes AST, CFG, and DFG information to expose control and data dependencies; a Knowledge Path that retrieves relevant CWE-derived vulnerability patterns and examples through hybrid dense--sparse retrieval; and a Semantic Path that summarizes the functional behavior of the code before the final judgment. These contexts are integrated into a unified instruction to guide the LLM toward more reliable vulnerability reasoning. Experiments on the PrimeVul pair test set show that VulTriage achieves state-of-the-art performance, outperforming existing deep learning and LLM-based baselines on key pair-wise and classification metrics. Further ablation studies verify the effectiveness of each path, and additional experiments on the Kotlin dataset demonstrate the generalization ability of VulTriage under low-resource and class-imbalanced settings. Our code is available at https://github.com/vinsontang1/VulTriage

Large language models are increasingly used for vulnerability detection, yet their reliability under different prompt formulations remains uncharacterized. We present PromptAudit, a controlled evaluation framework that isolates prompt effects by fixing the dataset, decoding, and parsing while varying only the prompting strategy. Using five prompting strategies across five open-weight models on 1,000 CVEs (6,074 code samples spanning 16 programming languages), we evaluate accuracy, recall, abstention, coverage, and effective F1. We find that standard chain-of-thought prompting achieves the strongest overall operational performance, while few-shot prompting provides model-dependent benefits that are most pronounced for prompt-sensitive models. In contrast, adaptive chain-of-thought frequently suppresses recall and self-consistency induces excessive abstention, sharply reducing effective performance. These results show that vulnerability detection behavior is jointly determined by the model and the prompt, and that prompt sensitivity is a first-class system property that must be explicitly characterized in evaluation and deployment.