Mistral Large 4 Release - Cyber Refusal (JailbreakBench, StrongREJECT, AgentHarm): leaderboard
Mistral Large 4 open-weight release post (2026-10-29): average refusal rate on the cyber-related prompts of JailbreakBench, StrongREJECT and AgentHarm, as measured and reported by Mistral for six models. Higher means more harmful cyber requests refused; reasoning effort not stated. Average refusal rate (%).
Metric: Average Refusal Rate (%). Source: mistral.ai. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Mistral Large 4 | 95.3 | #106 |
| 2 | Kimi K3 | 93.6 | #35 |
| 3 | GLM-5.3 | 87.1 | #44 |
| 4 | Gemma 4 31B | 84.1 | #212 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mistral-large-4-release-cyber-refusal-jailbreakbench-strongreject-agentharm · How It Works · Data refreshed daily, snapshot 2026-11-05.