Artificial Analysis Cyber Index: leaderboard
Artificial Analysis Cyber Index v1.0 (September 2026): equal-weight average of AA's own runs of CWE-Bench-AA (pass@1), DeepsecBench-AA (median F2) and CyberGym-E2E-AA (pass@1) on its Stirrup agent harness, measuring defensive work with source access: finding, reproducing and patching vulnerabilities. AA scores declined tasks as zero, so models that declined more than 5% of any component's tasks have no index cell here, and unmitigated trusted-access variants are excluded. Index (0-100).
Metric: Cyber Index (0-100). Source: artificialanalysis.ai. Saturation forecast: Around May 2028. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K3.1 Pro | 64.07 | #17 |
| 2 | DeepSeek V4.1 Flash (1017) (Max) | 56.81 | #27 (DeepSeek V4.1 Flash (1017)) |
| 3 | Grok 4.7 (xHigh) | 56.43 | #26 (Grok 4.7) |
| 4 | MiMo-V2.6-Pro | 56.07 | #48 |
| 5 | GPT-6 Luna (Max) | 52.69 | #61 (GPT-6 Luna) |
| 6 | mistral-large-4-oss (High) | 50.89 | #60 (mistral-large-4-oss) |
| 7 | GLM-5.3 Flash | 49.92 | #77 |
| 8 | Mistral Large 4 | 49.51 | #106 |
| 9 | Step 5 Preview | 47.71 | #53 |
| 10 | Muse Spark 1.3 (xHigh) | 44.3 | #31 (Muse Spark 1.3) |
| 11 | Kimi K3 (Max) | 41.27 | #35 (Kimi K3) |
| 12 | DeepSeek V4.1 Flash (Max) | 40.81 | #37 (DeepSeek V4.1 Flash) |
| 13 | GLM-5.3 (Max) | 36.38 | #44 (GLM-5.3) |
| 14 | Beam 1.0 | 33.7 | #102 |
| 15 | MiniMax-M3 | 22.39 | #151 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=artificial-analysis-cyber-index · How It Works · Data refreshed daily, snapshot 2026-11-05.