Agentic benchmark
CyberGym leaderboard
Every model the catalog carries a published CyberGym value for, ranked by that value.
CategoryAgentic
MeasureVulnerability reproduction and PoC generation
Tasks1,507 vulnerability analysis instances
DifficultyReal-world cybersecurity
A cybersecurity task benchmark for evaluating defensive cyber workflows and vulnerability-oriented agent performance.
CyberGym ranking
24 models with a published CyberGym value, ordered by that value, highest first. Models the source has not scored on this benchmark are not listed — they are unmeasured, not last.
| Rank | Model | Provider | Vulnerability reproduction and PoC generation |
|---|---|---|---|
| 1 |  DeepSeek V4.1 Flash | DeepSeek | 88.1 |
| 2 | SAFugu Cyber | Sakana AI | 86.9 |
| 3 |  Atria Dawn Preview | Shanghai Artificial Intelligence Laboratory | 86.5 |
| 4 |  Gemini 3.8 Flash Cyber | 86.2 | |
| 5 |  GLM-5.3 | Z.AI | 84.5 |
| 5 |  GPT-5.6 Sol | OpenAI | 84.5 |
| 7 |  Claude Mythos 5 | Anthropic | 83.8 |
| 8 | DeepSeek V4 Pro 0813 | DeepSeek | 83.3 |
| 9 | Gemini 3.5 Flash Cyber | 83.2 | |
| 10 | Claude Mythos Preview | Anthropic | 83.1 |
| 11 | GPT-5.5 | OpenAI | 81.8 |
| 11 | GPT-5.6 Terra | OpenAI | 81.8 |
| 13 | GPT-5.4 | OpenAI | 79.0 |
| 14 |  Hy4 preview | Tencent | 78.4 |
| 15 | GPT-5.6 Luna | OpenAI | 77.9 |
| 16 | DeepSeek V4 Flash 0731 | DeepSeek | 76.7 |
| 17 | Claude Opus 4.7 (Adaptive) | Anthropic | 73.1 |
| 18 | GLM-5.1 | Z.AI | 68.7 |
| 19 | Claude Opus 4.6 | Anthropic | 66.6 |
| 20 | Claude Sonnet 4.6 | Anthropic | 65.2 |
| 21 |  Muse Spark 1.1 | Meta | 59.0 |
| 22 | Claude Opus 4.5 | Anthropic | 50.6 |
| 23 | Muse Spark | Meta | 43.5 |
| 24 | GLM-5 | Z.AI | 43.2 |
Evidence key: Observed
Rows are ordered by the value CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at Scale published, highest first. The source does not state whether a higher value is the better result, so this page does not either: for a benchmark that measures a rate of failure, read the table from the bottom.