N-Day-Bench Benchmark Evaluates LLM Ability to Discover Real-World Vulnerabilities

Summary

N-Day-Bench is an adaptive benchmark designed to measure the capability of frontier language models to discover real-world software vulnerabilities (known as "N-days") that were disclosed after each model's knowledge cutoff date. The benchmark provides all models with an identical harness and context, aiming to prevent reward hacking and produce a fair comparison of genuine cybersecurity vulnerability discovery capabilities. Models are evaluated via OpenRouter-backed infrastructure, and the test