Refusal of harmful cyber requests (JailbreakBench, StrongREJECT, AgentHarm) (Cyber Misuse Refusal)
We show this table for reference; we do not rank on it.
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.
Benchmark score on Cyber Misuse Refusal — October 29, 2026
We compile the Cyber Misuse Refusal rows from provider self-reports and secondary reports. Mistral Large 4 leads the table at 96.6%, followed by Kimi K3 (94.0%) and GLM-5.3 (87.0%). We do not use these results to rank models overall.
Mistral Large 4
Mistral
Kimi K3
Moonshot AI
GLM-5.3
Z.AI
5 modelsAgenticCurrentDisplay onlyUpdated October 29, 2026
Benchmark score table (5 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Mistral Large 4Mistral | 96.6% | Not reported | Open |
| 2 | Kimi K3Moonshot AI | 94.0% | Not reported | Open |
| 3 | GLM-5.3Z.AI | 87.0% | Not reported | Open |
| 4 | Gemma 4 31BGoogle | 84.0% | Not reported | Open |
| 5 | DeepSeek V4 ProDeepSeek | 51.0% | Not reported | Open |
Among the reported Cyber Misuse Refusal rows, Mistral Large 4 is first at 96.6%. The third row is 9.6 points behind. The broader top-10 range is 45.6 points, so the table still separates the published systems.
5 models have been evaluated on Cyber Misuse Refusal. The benchmark falls in the Agentic category. Cyber Misuse Refusal is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About Cyber Misuse Refusal
Year
2026
Tasks
Cyber-related harmful prompts from three safety benchmarks
Format
Average refusal rate
Difficulty
Cyber misuse safety
Mistral reported these results in its October 29, 2026 Mistral Large 4 open-weight post, including the rival-model values in its chart. They are provider-run, so BenchLM shows them display-only and keeps them out of weighted scores. This is Mistral's own composite, not a public leaderboard; rival values were published as whole percentages.
Freshness and provenance
Version
Cyber Misuse Refusal 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does Cyber Misuse Refusal measure?
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.
Which model scores highest on Cyber Misuse Refusal?
Mistral Large 4 by Mistral currently leads with a score of 96.6% on Cyber Misuse Refusal.
How many models are evaluated on Cyber Misuse Refusal?
5 AI models have published results on Cyber Misuse Refusal in the BenchLM catalog.
Compare top models on Cyber Misuse Refusal
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 6,100+ subscribers.
One email each week. Unsubscribe anytime.