Benchmarking 13 AI models on rediscovering known CVEs

This benchmark evaluates 13 LLMs against 26 known CVEs using an AI code-analysis harness: GPT-5.6 achieved the highest recall (23/26), but cheaper mid-tier models can match results by pooling multiple runs, and open-weight models (GLM-5.2, Kimi K3) are rapidly improving—overall the report focuses on detection capability, cost-effectiveness, and variance rather than active exploitation.