N-Day-Bench – Can LLMs find real vulnerabilities in real codebases? (ndaybench.winfunc.com)

N-Day-Bench has been introduced to evaluate the ability of advanced language models (LLMs) to identify real-world vulnerabilities in software that emerged after their training cut-off dates. Developed by Winfunc Research, this benchmark aims to enhance cybersecurity by rigorously testing LLMs' capabilities in "vulnerability discovery" without allowing any manipulation of results. The benchmark is adaptive, meaning it updates test cases and model versions monthly, ensuring ongoing relevance in the evolving cybersecurity landscape. The results and traces from each benchmark run are publicly accessible, fostering transparency and collaboration within the AI/ML community. The latest benchmark run yielded average scores for models, with OpenAI's GPT-5.4 leading at 83.93, followed by Z-AI's GLM-5.1 and Anthropic's Claude-opus-4.6. The focus on real vulnerabilities signifies a pivotal shift in how LLMs can impact security practices, potentially paving the way for more robust coding standards and automated security assessments. The significance of N-Day-Bench lies in its potential to elevate the dialogue around machine learning applications in cybersecurity, challenging developers to refine these models for stronger protective measures against emerging threats.