benchgap
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 modelExploitBenchStealthMarkSource
GPT-6.2 Astra100.0%68.6%estimated ± 6.1 pp, medium confidence
GPT-6.2 Sol100.0%68.6%estimated ± 6.1 pp, medium confidence
GPT-6 Cyber100.0%68.6%estimated ± 6.1 pp, medium confidence
Claude Mythos 5.295.0%65.5%estimated ± 6.1 pp, medium confidence
Claude Mythos 578.0%56.5%estimated ± 6.1 pp, medium confidence
Claude Mythos Preview69.0%52.8%estimated ± 6.1 pp, medium confidence
SWE-2 Flash34.8%46.5%estimated ± 6.1 pp, medium confidence

Fitted on

ModelExploitBenchStealthMark
GLM-5.354.4%52.9%
GPT-5.6 Luna33.2%37.8%
GPT-5.6 Sol73.5%53.7%
GPT-5.6 Terra52.9%45.8%
GPT-6.1 Sol99.7%67.8%
GPT-6.1 Terra97.1%60.4%
GPT-6 Astra100.0%74.6%
Kimi K3.1 Pro74.7%56.7%
Kimi K332.0%50.6%
MiMo-V2.6-Flash25.3%49.3%
MiMo-V2.6-Pro47.9%47.9%