benchgap
Calibration

ExploitGym → CWE-bench v1

CWE-bench v1 is estimated from ExploitGym with a linear curve fitted on 6 models measured on both: y = 0.5166·x + 0.4576, R² = 0.83, cross-validated error 4.5 pp. It is used for 17 estimates.

Estimated modelExploitGymCWE-bench v1Source
GPT-6 Cyber48.5%70.8%estimated ± 4.5 pp, medium confidence
Claude Mythos 5.238.1%65.4%estimated ± 4.5 pp, medium confidence
GPT-6.1 Sol35.1%63.9%estimated ± 4.5 pp, medium confidence
GPT-5.6 Sol33.7%63.2%estimated ± 4.5 pp, medium confidence
GPT-5.6 Cyber32.1%62.3%estimated ± 4.5 pp, medium confidence
Claude Mythos 5.123.4%57.8%estimated ± 4.5 pp, medium confidence
GPT-5.6 Terra23.2%57.7%estimated ± 4.5 pp, medium confidence
Kimi K3.1 Pro20.7%56.4%estimated ± 4.5 pp, medium confidence
MiMo-V2.6-Pro17.8%55.0%estimated ± 4.5 pp, medium confidence
Claude Mythos Preview17.5%54.8%estimated ± 4.5 pp, medium confidence
GLM-5.315.0%53.5%estimated ± 4.5 pp, low confidence
GPT-5.513.4%52.7%estimated ± 4.5 pp, low confidence
GPT-5.6 Luna12.4%52.2%estimated ± 4.5 pp, low confidence
GPT-6 Luna11.6%51.7%estimated ± 4.5 pp, low confidence
GPT-5.46.0%48.9%estimated ± 4.5 pp, low confidence
MiMo-V2.6-Flash6.0%48.9%estimated ± 4.5 pp, low confidence
Muse Spark 1.10.8%46.2%estimated ± 4.5 pp, low confidence

Fitted on

ModelExploitGymCWE-bench v1
Claude Opus 5.533.3%67.0%
DeepSeek V4.1 Flash 101715.3%55.0%
Gemini 4 Argon39.5%68.0%
Gemini 4 Carbon53.5%71.0%
GPT-6 Astra42.4%68.0%
GPT-6 Sol22.1%52.0%