Agents & computer use

CyberGym

Cybersecurity benchmark (UC Berkeley Sunblaze) evaluating AI agents on real-world vulnerability discovery and exploitation tasks; scores are agent-system results that vary with harness and environment.

How models are tested

Scores are published numbers from the sources below. They are only comparable when the evaluation version, tools, agent harness, and sampling setup match. Missing scores are omitted rather than treated as zero.

Category: Agents & computer use
Metric: Percent
Direction: Higher is better
Coverage: 8 / 318Models in this catalog with a published score
Updated: 2026-09-29Catalog snapshot date
Source: github.com/sunblaze-ucb/cybergym

Scores

1
Atria Dawn Preview
Shanghai AI Labopen
86.5%
2
Gemini 3.8 Flash Cyber
Googleclosed
86.2%
3
GLM-5.3
Zhipu AIopen
84.5%
4
Claude Mythos 5
Anthropicclosed
83.8%
5
DeepSeek V4 Pro 0813
DeepSeekclosed
83.3%
6
DeepSeek V4 Flash 0731
DeepSeekopen
76.7%
7
DeepSeek V4 Pro
DeepSeekopen
52.7%
8
DeepSeek V4 Flash
DeepSeekopen
38.7%

Back to all benchmarks