CyberGym
Agents · 2025-06-03
CyberGym Level 1 measures whether an AI agent can reproduce a target real-world vulnerability by generating a working proof of concept from its description and unpatched codebase. The official leaderboard's default one-trial setting covers 1,507 historical vulnerabilities from 188 open-source projects.
Top models (higher is better)
| Model | Score |
|---|---|
| Opus 4.6 | 89.6 |
| GLM-5.1 | 84.0 |
| Claude Mythos 5 | 83.8 |
| GPT-5.6 Sol | 83.6 |
| Gemini 3.5 Flash Cyber | 83.2 |
| Claude Mythos Preview | 83.1 |
| GPT-5.5 | 81.8 |
| GPT-5.4 | 79.0 |
| MiniMax M3 | 73.1 |
| Sonnet 4.6 | 65.2 |
| GPT-5 | 60.2 |
| Muse Spark 1.1 | 59.0 |
| DeepSeek-V4-Pro | 57.7 |
| Opus 4.5 | 50.6 |
| Muse Spark | 43.5 |
| GLM-5 | 43.2 |
| Kimi K2.5 | 41.3 |
| Gemini 3.1 Pro Preview | 38.8 |
| Sonnet 4.5 | 28.9 |
| Opus 4.1 | 25.0 |
| GLM-4.7 | 23.5 |
| Sonnet 4 | 22.6 |
| Claude 3.7 Sonnet | 14.5 |
| GPT-4.1 | 9.4 |
| Gemini 2.5 Flash Preview 04-17 | 4.8 |
| DeepSeek-V3 | 3.6 |
| O4 Mini | 2.5 |
| Qwen3-235B-A22B | 1.9 |