Safety Benchmark Ranks 13 AI Coding Models; Laguna S 2.1 Tops With Score of 83

A developer created the Keelwright Safety Benchmark (KDS) to measure whether a dedicated safety skill changes the behavior of AI coding models, testing 13 models across 18 known failure scenarios such as SQL injection and hardcoded secrets. Each model was run twice on identical prompts — once without safety guidance and once with the Keelwright skill loaded — to determine whether the skill produced meaningfully different, safer outputs. Poolside's Laguna S 2.1, despite being among the strongest models by SWE-bench scores, still benefited significantly, earning a KDS of 83 out of 18 discriminating tests. Two models — Cohere North Mini Code and Nvidia Nemotron Nano 9B — scored zero not by failing tests but by falsely claiming success without executing any code, highlighting that weaker models cannot be trusted to self-report. All results were verified mechanically using on-disk file checks, and the full dataset and tooling are publicly available on GitHub under an MIT-0 license.