Calibration
SEC-Bench Pro → ExploitBench
ExploitBench is estimated from SEC-Bench Pro with a Hill curve fitted on 5 models measured on both: y = 0.1074 + (1.2000 − 0.1074)·x^6.00 / (0.68048^6.00 + x^6.00), R² = 0.92, cross-validated error 12.2 pp. It is used for 4 estimates.
| Estimated model | SEC-Bench Pro | ExploitBench | Source |
|---|---|---|---|
| DeepSeek V4.1 Flash | 62.8% | 52.5% | estimated ± 12.2 pp, low confidence |
| GPT-5.5 | 45.8% | 20.0% | estimated ± 12.2 pp, low confidence |
| GPT-6 Luna | 34.2% | 12.5% | estimated ± 12.2 pp, low confidence |
| GPT-6 Sol | 66.3% | 61.1% | estimated ± 12.2 pp, low confidence |