Real-world defensive-security benchmarks expose model ranking trade-offs
GreyNoise’s AI Cyber Index evaluates 46 models on production-derived attack data and reports capability, refusals, time, and cost separately. Opus 5.5 reportedly leads overall, while Grok 4.5 tops a refusal-penalized ranking, illustrating how scoring choices change results. The benchmark is promising but creator-reported; evaluation details and independent validation remain important.
Sources (2)
Updated Nov 16, 2026