Benchmarks / Cyber
DeepsecBench-AA
DeepsecBench-AA (based on Vercel's DeepsecBench), part of the AA Cyber Index: finding and fixing vulnerabilities in real codebases; score.
DeepsecBench-AA
Higher is better · best effort per model
full table below
| # | Model | Effort | Score |
|---|---|---|---|
| 1 | GPT-6 Sol | max | 48.1% |
| 2 | GPT-6 Astra | max | 36.9% |
| 3 | DeepSeek V4.1 Flash | max | 31.2% |
| 4 | Grok 4.7 | xhigh | 26.9% |
| 5 | Opus 5.5 | max | 26.9% |
| 6 | MiMo V2.6 Pro | – | 26.2% |
| 7 | Opus 5 | max | 25.9% |
| 8 | GPT-5.6 Terra | max | 25.1% |
| 9 | GLM 5.3 | max | 24.3% |
| 10 | GPT-6 Luna | max | 23.5% |
| 11 | Fable 5.1 | max | 23.5% |
| 12 | GLM 5.3 Flash | max | 19.9% |
| 13 | Mistral Large 4 | – | 16.0% |
| 14 | Step 5 Preview | – | 15.8% |
| 15 | Gemini 3.8 Flash | high | 14.5% |
| 16 | Muse Spark 1.3 | xhigh | 12.5% |
| 17 | Kimi K3 | max | 12.5% |
| 18 | Qwen 3.8 27B | xhigh | 11.5% |
| 19 | Qwen 3.8 2.4T-A95B | – | 9.5% |
| 20 | Beam 1.0 | – | 6.9% |
| 21 | MiniMax M3 | – | 1.1% |
| 22 | Nemotron 3 Ultra | max | 1.0% |