# CyberGym

Gives AI agents a description of a real security bug in open-source software plus the code, and asks them to write an input that actually triggers the bug. It shows how good AI has become at the hands-on work of finding and reproducing vulnerabilities, which helps defenders and could help attackers.

Category: safety. Type: leaderboard. Status: active.
Metric: Share of 1,507 real vulnerabilities the agent successfully reproduced (%) (higher is better).
Source: CyberGym team (UC Berkeley) — https://www.cybergym.io/assets/data/cybergym.json, results as of 2026-10-24, retrieved 2026-11-17, license: not stated (benchmark code Apache-2.0).

| Rank | Model | Agent | Success Rate | Date | Source |
|---|---|---|---|---|---|
| 1 | Creation Model | Claude Code | 90.6% | 2026-10-21 | Creation Lab (Creation Lab, Ant Group) |
| 2 | XekRung-1.5-27B-Preview (finetuned from Qwen3.8-27B) | XekRung Agent | 88.9% | 2026-09-13 | Alibaba Security |
| 3 | Gemini 3.8 Flash Cyber | Google Antigravity | 86.3% | 2026-09-10 | Google |
| 4 | GPT-5.5-Cyber | OpenAI Agent | 85.6% | 2026-06-22 | OpenAI |
| 5 | GLM-5.3 | Claude Code | 84.5% | 2026-08-14 | Zhipu AI |
| 6 | Feyospace-v1.1 (finetuned from Qwen3.8-27B) | Claude Code | 84.4% | 2026-09-23 | Vera Praxis Lab |
| 7 | DeepSeek-V4-Pro | DeepSeek Agent | 83.3% | 2026-08-13 | DeepSeek |
| 8 | Claude Mythos Preview | Anthropic Agent | 83.1% | 2026-04-07 | Anthropic |
| 9 | GPT-5.5 | OpenAI Agent | 81.8% | 2026-04-23 | OpenAI |
| 10 | Basin (basin1-0912-pre-dpo) | Codex CLI | 80.0% | 2026-09-23 | Submersion AI |
| 11 | Grok 4.6 | Grok Build | 79.7% | 2026-08-12 | xAI |
| 12 | Grok 4.5 | Grok Build | 79.0% | 2026-07-14 | xAI |
| 13 | GPT-5.4 | OpenAI Agent | 79.0% | 2026-04-23 | OpenAI |
| 14 | GLM-5.2 | Claude Code | 77.2% | 2026-07-21 | Zhipu AI |
| 15 | DeepSeek-V4-Flash | DeepSeek Agent | 76.7% | 2026-07-31 | DeepSeek |
| 16 | GLM-5.1 | Claude Code | 68.7% | 2026-04-07 | Zhipu AI |
| 17 | Claude Opus 4.6 | Anthropic Agent | 66.6% | 2026-02-05 | Anthropic |
| 18 | GPT-5.4 | Codex CLI | 66.3% | 2026-04-07 | Zhipu AI |
| 19 | Claude Sonnet 4.6 | Anthropic Agent | 65.2% | 2026-02-17 | Anthropic |
| 20 | Muse Spark 1.1 (helpful-only) | Meta Agent | 59.0% | 2026-07-09 | Meta AI |
| 21 | DeepSeek-V4-Pro Preview | Claude Code | 57.7% | 2026-07-13 | XDxAI |
| 22 | Claude Opus 4.5 | Anthropic Agent | 50.6% | 2026-02-05 | Anthropic |
| 23 | Muse Spark | Meta Agent | 43.5% | 2026-04-14 | Meta AI |
| 24 | GLM-5 | Claude Code | 43.2% | 2026-02-12 | Zhipu AI |
| 25 | Kimi K2.5 | Kimi Agent | 41.3% | 2026-02-02 | Kimi |

Notes: Level 1 (vulnerability description + codebase), single-trial runs. Reflects the site's default 'Model-focused' view, which excludes agent-focused systems and runs using the 'dynamic' (cleaned vulnerable Docker image) or 'test-time memory' settings; in the 'All' view many agent systems score 92–99%. Scores are mostly submitter-reported (Source column). The site shows entries above 90% as unranked cards in random order. as_of = newest entry date in the data file.
