Skip to content

CyberGym

A cybersecurity task benchmark for evaluating defensive cyber workflows and vulnerability-oriented agent performance.

25 models

Higher is better.

  1. 1MiMo-V2.6-FlashXiaomi95.1
  2. 2MiMo-V2.6-ProXiaomi94.0
  3. 3DeepSeek V4.1 Flash 1017DeepSeek88.1
  4. 4Ling 3.1 FlashInclusionAI87.9
  5. 5Atria Dawn PreviewShanghai Artificial Intelligence Laboratory86.5
  6. 6Step 5 PreviewStepFun84.7
  7. 7GPT-5.6 SolOpenAI84.5
  8. 8GLM-5.3Z.AI84.5
  9. 9Claude Mythos 5Anthropic83.8
  10. 10DeepSeek V4 Pro 0813DeepSeek83.3
  11. 11MiniMax M3.1 FlashMiniMax82.2
  12. 12GPT-5.6 TerraOpenAI81.8
  13. 13GPT-5.5OpenAI81.8
  14. 14GPT-5.4OpenAI79.0
  15. 15Hy4 previewTencent78.4
  16. 16GPT-5.6 LunaOpenAI77.9
  17. 17DeepSeek V4 Flash 0731DeepSeek76.7
  18. 18Claude Opus 4.7 (Adaptive)Anthropic73.1
  19. 19GLM-5.1Z.AI68.7
  20. 20Claude Opus 4.6Anthropic66.6
  21. 21Claude Sonnet 4.6Anthropic65.2
  22. 22Muse Spark 1.1Meta59.0
  23. 23Claude Opus 4.5Anthropic50.6
  24. 24Muse SparkMeta43.5
  25. 25GLM-5Z.AI43.2
Higher is better. Bar colour: blue proprietary, orange open weights, green other.

About this benchmark

Metric
Vulnerability reproduction and PoC generation
Tasks
1,507 vulnerability analysis instances
Authors
Zhun Wang, Tianneng Shi, Jingxuan He, Matthew Cai, Jialin Zhang, Dawn Song (2026)

Scores are as reported by model developers or evaluators and may use different settings (for example reasoning effort or tools). Compare with care.