# Artificial Analysis CyberGym-E2E-AA (CyberGym-E2E-AA)

> Artificial Analysis' implementation of Berkeley RDI's CyberGym-E2E: agents find a real memory-safety vulnerability in a C/C++ open-source project, write a proof-of-concept input that crashes it, and patch it without breaking the project's tests.

Canonical page: https://benchlm.ai/benchmarks/aacybergyme2e

- Category: [Agentic](/agentic)
- Last updated: October 29, 2026

## About CyberGym-E2E-AA

- Year: 2026
- Tasks: 131 end-to-end vulnerability reproduction and patching tasks
- Format: Pass@1
- Difficulty: Real-world cybersecurity
- Paper: [CyberGym-E2E-AA Benchmark Leaderboard](https://artificialanalysis.ai/evaluations/cybergym-e2e-aa)

131 tasks, one per OSS-Fuzz project, 90 minutes each in the Stirrup harness. Separate from the provider-reported CyberGym rows. Independently run by Artificial Analysis under its published harness. Display-only; not yet admitted as independent-run evidence.

CyberGym-E2E-AA is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (19 models)

| Rank | Model | Creator | Score |
|------|-------|---------|-------|
| 1 | [Kimi K3.1 Pro](/models/kimi-3-1-pro) | Moonshot AI | 94.7% |
| 2 | [Mistral Large 4](/models/mistral-large-4) | Mistral | 81.7% |
| 3 | [MiMo-V2.6-Pro](/models/mimo-v2-6-pro) | Xiaomi | 78.6% |
| 4 | [GPT-6 Luna](/models/gpt-6-luna) | OpenAI | 77.9% |
| 5 | [Step 5 Preview](/models/step-5-preview) | StepFun | 74.8% |
| 6 | [GLM-5.3-Flash](/models/glm-5-3-flash) | Z.AI | 74.0% |
| 7 | [Grok 4.7](/models/grok-4-7) | xAI | 74.0% |
| 8 | [Kimi K3](/models/kimi-k3) | Moonshot AI | 58.0% |
| 9 | [Beam 1.0 501B-A23B](/models/reflection-beam) | Reflection AI | 45.0% |
| 10 | [GLM-5.3](/models/glm-5-3) | Z.AI | 29.0% |
| 11 | [MiniMax M3](/models/minimax-m3) | MiniMax | 24.4% |
| 12 | [DeepSeek V4.1 Flash 1017](/models/deepseek-v4-1-flash) | DeepSeek | 22.9% |
| 13 | [Nemotron 3 Ultra](/models/nemotron-3-ultra) | NVIDIA | 10.7% |
| 14 | [Gemini 3.8 Flash](/models/gemini-3-8-flash) | Google | 8.4% |
| 15 | [Claude Opus 5.5](/models/claude-opus-5-5) | Anthropic | 0.8% |
| 16 | [Claude Fable 5.1](/models/claude-fable-5-1) | Anthropic | 0.8% |
| 17 | [GPT-6 Astra](/models/gpt-6-astra) | OpenAI | 0.0% |
| 18 | [GPT-6 Sol](/models/gpt-6-sol) | OpenAI | 0.0% |
| 19 | [Qwen3.8-27B](/models/qwen3-8-27b) | Alibaba | 0.0% |

## FAQ

### What does CyberGym-E2E-AA measure?

Artificial Analysis' implementation of Berkeley RDI's CyberGym-E2E: agents find a real memory-safety vulnerability in a C/C++ open-source project, write a proof-of-concept input that crashes it, and patch it without breaking the project's tests.

### Which model scores highest on CyberGym-E2E-AA?

Kimi K3.1 Pro by Moonshot AI currently leads with a score of 94.7% on CyberGym-E2E-AA.

### How many models are evaluated on CyberGym-E2E-AA?

19 AI models have published results on CyberGym-E2E-AA in the BenchLM catalog.

### Does CyberGym-E2E-AA affect BenchLM's overall score?

Not directly. CyberGym-E2E-AA is still displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Compare Top Models on CyberGym-E2E-AA

- [Kimi K3.1 Pro vs Mistral Large 4](/compare/kimi-3-1-pro-vs-mistral-large-4)
- [Mistral Large 4 vs MiMo-V2.6-Pro](/compare/mimo-v2-6-pro-vs-mistral-large-4)
- [MiMo-V2.6-Pro vs GPT-6 Luna](/compare/gpt-6-luna-vs-mimo-v2-6-pro)
- [GPT-6 Luna vs Step 5 Preview](/compare/gpt-6-luna-vs-step-5-preview)