A new benchmark called N-Day-Bench evaluates the ability of large language models to identify real-world vulnerabilities within actual codebases. The framework utilizes strict GitHub advisories and OpenRouter-backed models to rigorously test AI-driven security scanning capabilities.