Calibration
ExploitBench → StealthMark
StealthMark is estimated from ExploitBench with a Hill curve fitted on 11 models measured on both: y = 0.4603 + (1.2000 − 0.4603)·x^3.94 / (1.23275^3.94 + x^3.94), R² = 0.81, cross-validated error 6.1 pp. It is used for 7 estimates.
| Estimated model | ExploitBench | StealthMark | Source |
|---|---|---|---|
| GPT-6.2 Astra | 100.0% | 68.6% | estimated ± 6.1 pp, medium confidence |
| GPT-6.2 Sol | 100.0% | 68.6% | estimated ± 6.1 pp, medium confidence |
| GPT-6 Cyber | 100.0% | 68.6% | estimated ± 6.1 pp, medium confidence |
| Claude Mythos 5.2 | 95.0% | 65.5% | estimated ± 6.1 pp, medium confidence |
| Claude Mythos 5 | 78.0% | 56.5% | estimated ± 6.1 pp, medium confidence |
| Claude Mythos Preview | 69.0% | 52.8% | estimated ± 6.1 pp, medium confidence |
| SWE-2 Flash | 34.8% | 46.5% | estimated ± 6.1 pp, medium confidence |
Fitted on
| Model | ExploitBench | StealthMark |
|---|---|---|
| GLM-5.3 | 54.4% | 52.9% |
| GPT-5.6 Luna | 33.2% | 37.8% |
| GPT-5.6 Sol | 73.5% | 53.7% |
| GPT-5.6 Terra | 52.9% | 45.8% |
| GPT-6.1 Sol | 99.7% | 67.8% |
| GPT-6.1 Terra | 97.1% | 60.4% |
| GPT-6 Astra | 100.0% | 74.6% |
| Kimi K3.1 Pro | 74.7% | 56.7% |
| Kimi K3 | 32.0% | 50.6% |
| MiMo-V2.6-Flash | 25.3% | 49.3% |
| MiMo-V2.6-Pro | 47.9% | 47.9% |