# Cyber Misuse Refusal leaderboard

Agentic benchmark

> Every figure below is reproduced as its upstream source published it: nothing is modelled, estimated, interpolated or converted. `Unavailable` means no source published the value — it is never a zero. Each value carries its evidence state and the date it was observed.

Page: https://modelchoice.dev/leaderboards/benchmarks/agentic-cyber-harmful-refusal

Refusal of harmful cyber requests (JailbreakBench, StrongREJECT, AgentHarm). Every model the catalog carries a published Cyber Misuse Refusal value for, ranked by that value.

## Benchmark record

| Field | Value |
| --- | --- |
| Category | Agentic |
| Measure | Average refusal rate |
| Tasks | Cyber-related harmful prompts from three safety benchmarks |
| Difficulty | Cyber misuse safety |
| Published by | [Mistral Large 4 is now open weight](https://mistral.ai/news/mistral-large-4-oss/) |

Average refusal rate on the cyber-related prompts of JailbreakBench, StrongREJECT and AgentHarm, as measured by Mistral. Higher means the model refuses more malicious cyber requests.

## Cyber Misuse Refusal ranking

4 models with a published Cyber Misuse Refusal value, ordered by that value, highest first. Models the source has not scored on this benchmark are not listed — they are unmeasured, not last.

| Rank | Model | Provider | Average refusal rate | Evidence |
| --- | --- | --- | --- | --- |
| 1 | [Mistral Large 4](https://modelchoice.dev/models/mistral-large-4) | Mistral | 96.6 | Observed 2026-11-21 · source benchlm:models |
| 2 | [Kimi K3](https://modelchoice.dev/models/kimi-3) | Moonshot AI | 94.0 | Observed 2026-11-21 · source benchlm:models |
| 3 | [GLM-5.3](https://modelchoice.dev/models/glm-5-3) | Z.AI | 87.0 | Observed 2026-11-21 · source benchlm:models |
| 4 | [Gemma 4 31B](https://modelchoice.dev/models/gemma-4-31b) | Google | 84.0 | Observed 2026-11-21 · source benchlm:models |

Rows are ordered by the value the source published, highest first. The source does not state whether a higher value is the better result, so this document does not either: for a benchmark that measures a rate of failure, read the table from the bottom.
