Calibration
SEC-Bench Pro → ExploitBench
ExploitBench is estimated from SEC-Bench Pro with a Hill curve fitted on 6 models measured on both: y = 0.1038 + (1.2000 − 0.1038)·x^6.00 / (0.67821^6.00 + x^6.00), R² = 0.93, cross-validated error 10.7 pp. It is used for 6 estimates.
| Estimated model | SEC-Bench Pro | ExploitBench | Source |
|---|---|---|---|
| Gemini 4 Carbon | 88.2% | 100.0% | estimated ± 10.7 pp, low confidence |
| Gemini 4 Argon | 83.4% | 95.4% | estimated ± 10.7 pp, low confidence |
| GPT-6 Sol | 66.3% | 61.5% | estimated ± 10.7 pp, low confidence |
| DeepSeek V4.1 Flash 1017 | 62.8% | 52.8% | estimated ± 10.7 pp, low confidence |
| GPT-5.5 | 45.8% | 19.9% | estimated ± 10.7 pp, low confidence |
| GPT-6 Luna | 34.2% | 12.2% | estimated ± 10.7 pp, low confidence |
Fitted on
| Model | SEC-Bench Pro | ExploitBench |
|---|---|---|
| GPT-5.6 Sol | 71.2% | 73.5% |
| GPT-6.1 Sol | 78.8% | 99.7% |
| GPT-6.1 Terra | 83.3% | 97.1% |
| GPT-6 Astra | 85.4% | 100.0% |
| MiMo-V2.6-Flash | 47.5% | 25.3% |
| MiMo-V2.6-Pro | 66.3% | 47.9% |