Part of the benchmark canon: 6 entries, all measure the model. How to read the Paladin column is defined on the canon page.
| Benchmark | Organization | Measures | Paladin |
|---|---|---|---|
| HarmBench (top 12) | Center for AI Safety | Standardized red-teaming with classifier grading | Measures the model |
| StrongREJECT | UC Berkeley | Jailbreak scoring that discounts low-quality answers | Measures the model |
| JailbreakBench | Academic | Open jailbreak robustness suite | Measures the model |
| XSTest | Academic (Rottger et al.) | Over-refusal of safe requests | Measures the model |
| SORRY-Bench | Princeton / Virginia Tech | Fine-grained refusal across 45 unsafe topics | Measures the model |
| AdvBench | CMU / Center for AI Safety | Adversarial harmful-behavior prompts | Measures the model |
Previous: Software security and code. Next: CBRN knowledge and bio uplift.