BlogLLM Benchmark Scores as Published, Side by Side
Cyber Refusal Rate: Results as Published
Every result the catalog holds for Cyber Refusal Rate: 6 models from 4 developers, read from 1 page. Each number is as printed by the company named beside it; nothing is ranked or rescored. Models are listed by developer, newest first.
Area Safety and AlignmentMaintainer Mistral AIIntroduced 2026Data version 2026103001, as of 2026-10-30
Mistral reports how often a model refuses cybersecurity prompts that it treats as malicious. The prompts are drawn from JailbreakBench, StrongREJECT, and AgentHarm. The score is the average refusal rate across those cyber prompts. The post does not print how many prompts are used or how the three sets are weighted.
Cyber refusal rate (average of JailbreakBench, StrongREJECT, AgentHarm)
| Model | Value | Printed by | Setup as printed |
|---|---|---|---|
| DeepSeek-V4-Pro-0813DeepSeek | 77.7% | Mistral AI, Mistral Large 4 (preview)checked 2026-10-15 | - |
| Kimi K3Moonshot AI, released 2026-07-16 | 93.7% | Mistral AI, Mistral Large 4 (preview)checked 2026-10-15 | - |
| Kimi K2.6Moonshot AI, released 2026-04-20 | 90.7% | Mistral AI, Mistral Large 4 (preview)checked 2026-10-15 | - |
| GLM-5.3Z.ai, released 2026-08-18 | 87% | Mistral AI, Mistral Large 4 (preview)checked 2026-10-15 | - |
| GLM-5.2Z.ai, released 2026-06-16 | 90% | Mistral AI, Mistral Large 4 (preview)checked 2026-10-15 | - |
| Mistral Large 4 (preview)Mistral AI, released 2026-10-06 | 95.3% | Mistral AI, Mistral Large 4 (preview)checked 2026-10-15 | - |
Every number here is read from the same catalog file as the benchmark table, data version 2026103001, and is the number the table shows for that model and test. Where one page printed a model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.