Calibration
ExploitGym → AA ITBench
AA ITBench is estimated from ExploitGym with a offset logistic curve fitted on 7 models measured on both: y = 0.4626 + (0.5240 − 0.4626) / (1 + exp(−105.19·(x − 0.2205))), R² = 0.63, cross-validated error 3.9 pp. It is used for 6 estimates.
| Estimated model | ExploitGym | AA ITBench | Source |
|---|---|---|---|
| Claude Mythos Preview | 17.5% | 46.3% | estimated ± 3.9 pp, medium confidence |
| GPT-5.6 Luna | 12.4% | 46.3% | estimated ± 3.9 pp, low confidence |
| GPT-6.1 Sol | 35.1% | 52.4% | estimated ± 3.9 pp, medium confidence |
| GPT-6 Luna | 11.6% | 46.3% | estimated ± 3.9 pp, low confidence |
| MiMo-V2.6-Flash | 6.0% | 46.3% | estimated ± 3.9 pp, low confidence |
| MiMo-V2.6-Pro | 17.8% | 46.3% | estimated ± 3.9 pp, medium confidence |