Calibration
ExploitBench → SEC-Bench Pro
SEC-Bench Pro is estimated from ExploitBench with a Michaelis–Menten curve fitted on 5 models measured on both: y = 1.0655·x / (0.31264 + x), R² = 0.95, cross-validated error 4.3 pp. It is used for 6 estimates.
| Estimated model | ExploitBench | SEC-Bench Pro | Source |
|---|---|---|---|
| Claude Mythos 5 | 78.0% | 76.1% | estimated ± 4.3 pp, medium confidence |
| Claude Mythos Preview | 69.0% | 73.3% | estimated ± 4.3 pp, medium confidence |
| GLM-5.3 | 54.4% | 67.7% | estimated ± 4.3 pp, medium confidence |
| GPT-5.6 Luna | 33.2% | 54.9% | estimated ± 4.3 pp, medium confidence |
| GPT-5.6 Terra | 52.9% | 67.0% | estimated ± 4.3 pp, medium confidence |
| Kimi K3 | 32.0% | 53.9% | estimated ± 4.3 pp, medium confidence |