Arastoo Zibaeirad

Arastoo Zibaeirad, Marco Vieira

Large Language Models (LLMs) have shown promise in tasks like code translation, prompting interest in their potential for automating software vulnerability detection (SVD) and patching (SVP). To further research in this area, establishing a benchmark is essential for evaluating the strengths and limitations of LLMs in these tasks. Despite their capabilities, questions remain regarding whether LLMs can accurately analyze complex vulnerabilities and generate appropriate patches. This paper introduces VulnLLMEval, a framework designed to assess the performance of LLMs in identifying and patching vulnerabilities in C code. Our study includes 307 real-world vulnerabilities extracted from the Linux kernel, creating a well-curated dataset that includes both vulnerable and patched code. This dataset, based on real-world code, provides a diverse and representative testbed for evaluating LLM performance in SVD and SVP tasks, offering a robust foundation for rigorous assessment. Our results reveal that LLMs often struggle with distinguishing between vulnerable and patched code. Furthermore, in SVP tasks, these models tend to oversimplify the code, producing solutions that may not be directly usable without further refinement.

Arastoo Zibaeirad, Marco Vieira

Large Language Models (LLMs) are increasingly being studied for Software Vulnerability Detection (SVD) and Repair (SVR). Individual LLMs have demonstrated code understanding abilities, but they frequently struggle when identifying complex vulnerabilities and generating fixes. This study presents DVDR-LLM, an ensemble framework that combines outputs from diverse LLMs to determine whether aggregating multiple models reduces error rates. Our evaluation reveals that DVDR-LLM achieves 10-12% higher detection accuracy compared to the average performance of individual models, with benefits increasing as code complexity grows. For multi-file vulnerabilities, the ensemble approach demonstrates significant improvements in recall (+18%) and F1 score (+11.8%) over individual models. However, the approach raises measurable trade-offs: reducing false positives in verification tasks while simultaneously increasing false negatives in detection tasks, requiring careful decision on the required level of agreement among the LLMs (threshold) for increased performance across different security contexts. Artifact: https://github.com/Erroristotle/DVDR_LLM

Arastoo Zibaeirad, Marco Vieira

Whether LLMs scoring well on vulnerability benchmarks genuinely reason about security or merely pattern-match on contaminated data remains unresolved. We present CWE-Trace, a framework for LLM vulnerability detection built from 834 manually curated Linux kernel samples spanning 74 CWEs. The framework enforces a strict temporal split (pre-2025 historical set / post-cutoff leakage-free set), preserves context-aware vulnerable--patched pairs, and introduces two diagnostic metrics: the Directional Failure Index (DFI) and Hierarchical Distance and Direction (HDD). We evaluate eight vanilla LLMs and 15 LoRA fine-tuned variants across non-targeted detection, targeted detection, and CWE classification. Our analysis yields two key results. First, data contamination provides no measurable advantage. Function-level analysis shows that 84% of nominally contaminated samples carry no usable memorization signal: vulnerable functions are absent or cross-mapped across datasets, and ~31% of contaminated samples carry CWE misclassification. Second, backbone directional priors dominate fine-tuning. Models exhibit stable, systematic failure modes (DFI ranging from -85.5 to +94.8 pp) that persist from historical to post-cutoff data and resist correction. Fine-tuning shifts the output threshold without changing the decision policy. This is calibration without comprehension: output distributions adapt to training data while the underlying security reasoning remains absent. The weakest backbone at binary detection (DeepSeek-R1) gains the most in coarse CWE classification, revealing that detection and understanding are decoupled capabilities. The best detection score reaches only 52.1% (+2.1 pp above chance); exact CWE ranking remains below 1.3% Top-1 accuracy, confirming that current LLMs lack reliable security reasoning for systems software, regardless of fine-tuning strategy.