Zhun Wang et al.

CyberGym

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

In index
Agentic
Source ↗

Models27

Top result95.1%MiMo-V2.6-Flash

Top-3 spread7.0 pts

Year2026

Result and price

Results

27 results

#ModelResult

01

MiMo-V2.6-Flash95.1%

02

MiMo-V2.6-Pro94.0%

03

DeepSeek V4.1 Flash88.1%

04SAFugu Cyber86.9%

05SAAtria Dawn Preview86.5%

06

Gemini 3.8 Flash Cyber86.2%

07STStep 5 Preview84.7%

08

GPT-5.6 Sol84.5%

08

GLM-5.384.5%

10

Claude Mythos 583.8%

11

DeepSeek V4 Pro 081383.3%

12

Gemini 3.5 Flash Cyber83.2%

13

Claude Mythos Preview83.1%

14

GPT-5.6 Terra81.8%

14

GPT-5.581.8%

16

GPT-5.479.0%

17TEHy4 preview78.4%

18

GPT-5.6 Luna77.9%

19

DeepSeek V4 Flash 073176.7%

20

Claude Opus 4.7 (Adaptive)73.1%

21

GLM-5.168.7%

22

Claude Opus 4.666.6%

23

Claude Sonnet 4.665.2%

24

Muse Spark 1.159.0%

25

Claude Opus 4.550.6%

26

Muse Spark43.5%

27

GLM-543.2%