Atlas

Benchmarks

← All benchmarks

SEC-Bench Pro - release-reported cohort

Code

Proof-of-concept generation on complex vulnerable software, as reported in model release tables. Exact cohort is not established as CAISI's 183-task subset; source task count remains unknown.

Top models (higher is better)

ModelScore
GPT-6 Astra85.4
GPT-5.6 Sol71.2
MiMo-V2.6-Pro66.3
DeepSeek V4.1 Flash62.8
GPT-5.6 Terra57.7
GPT-5.6 Luna48.9
MiMo-V2.6-Flash47.5
Loading Atlas data…