safety benchmark
CyberGym Leaderboard
CyberGym is a benchmark for evaluating AI agents on cybersecurity tasks in a controlled environment.
Updated Oct 23, 2026
Models23
Model coverage23
MetricScore
EvidenceB
CyberGym Ranking
Higher score ranks better on this benchmark.
Sort by
Rank | Model | Score | Percentile | Participants | Evidence | Evaluated |
|---|
| Rank01 | ModelXI | Score95.10% | Percentile100.00% | Participants23 | EvidenceC | Evaluated |
| Rank02 | ModelXI | Score94.00% | Percentile95.45% | Participants23 | EvidenceC | Evaluated |
| Rank03 | ModelDE | Score88.10% | Percentile90.91% | Participants23 | EvidenceC | Evaluated |
| Rank04 | ModelIN | Score87.90% | Percentile86.36% | Participants23 | EvidenceC | Evaluated |
| Rank05 | ModelSA | Score86.50% | Percentile81.82% | Participants23 | EvidenceC | Evaluated |
| Rank06 | ModelGO | Score86.20% | Percentile77.27% | Participants23 | EvidenceC | Evaluated |
| Rank07 | ModelZA | Score84.50% | Percentile72.73% | Participants23 | EvidenceC | Evaluated |
| Rank08 | ModelDE | Score83.30% | Percentile68.18% | Participants23 | EvidenceC | Evaluated |
| Rank09 | ModelGO | Score83.20% | Percentile63.64% | Participants23 | EvidenceC | Evaluated |
| Rank10 | ModelAN | Score83.10% | Percentile59.09% | Participants23 | EvidenceC | Evaluated |
| Rank11 | ModelMI | Score82.20% | Percentile54.55% | Participants23 | EvidenceC | Evaluated |
| Rank12 | ModelMA | Score82.00% | Percentile50.00% | Participants23 | EvidenceC | Evaluated |
| Rank13 | ModelOP | Score81.80% | Percentile45.45% | Participants23 | EvidenceC | Evaluated |
| Rank14 | ModelXA | Score80.30% | Percentile40.91% | Participants23 | EvidenceC | Evaluated |
| Rank15 | ModelAN | Score78.80% | Percentile36.36% | Participants23 | EvidenceC | Evaluated |
| Rank16 | ModelTE | Score78.40% | Percentile31.82% | Participants23 | EvidenceC | Evaluated |
| Rank17 | ModelDE | Score76.70% | Percentile27.27% | Participants23 | EvidenceC | Evaluated |
| Rank18 | ModelAN | Score73.80% | Percentile22.73% | Participants23 | EvidenceC | Evaluated |
| Rank19 | ModelAN | Score73.10% | Percentile18.18% | Participants23 | EvidenceC | Evaluated |
| Rank20 | ModelBY | Score70.20% | Percentile13.64% | Participants23 | EvidenceC | Evaluated |
| Rank21 | ModelZA | Score68.70% | Percentile9.09% | Participants23 | EvidenceC | Evaluated |
| Rank22 | ModelBY | Score67.00% | Percentile4.55% | Participants23 | EvidenceC | Evaluated |
| Rank23 | ModelMA | Score41.30% | Percentile0.00% | Participants23 | EvidenceC | Evaluated |
CyberGym Highlights
The leading models and scores on this benchmark.
CyberGym Score Distribution
A closer view of the leading scores on this benchmark.
The Top AI Models for CyberGym
The first five results on this benchmark, with official price and output speed added where the model identity can be matched.
What is CyberGym?
What CyberGym measures and how its scores work.
CyberGym is a safety benchmark for evaluating AI agents on cybersecurity tasks.
It measures AI agents' ability to identify vulnerabilities, perform security analysis, and complete security-related challenges, reported as a Score ratio.
Scores are shown in ratio. This benchmark is not independently verified and has an evidence level of B.
- Family
- CyberGym
- Modality
- text
- Primary category
- safety
- Score direction
- higher
- LLMBoard eligible
- No
- Evaluation key
- overall
LLM Stats. Benchmark scores retain their original unit. Overall score eligibility is shown separately.
FAQ
Common questions about CyberGym.
Which model scores highest on CyberGym?
MiMo-V2.6-Flash is currently ranked first with 95.10%.
What are the top three models on CyberGym?
The current leaders are MiMo-V2.6-Flash (95.10%), MiMo-V2.6-Pro (94.00%), and DeepSeek-V4.1-Flash (88.10%).
Which CyberGym model has the lowest official input price?
MiMo-V2.6-Flash has the lowest matched official input price at $0.14 input / $0.28 output per 1M tokens.
Which models are fastest among CyberGym results?
The fastest matched records are Claude Opus 4.8 (42.00 tok/s via Vertex AI), Claude Opus 4.6 (42.00 tok/s via Anthropic), and Claude Opus 4.7 (42.00 tok/s via Vertex AI).
Does the highest score result prove overall model quality?
No. This benchmark measures one defined capability or task. The overall LLMBoard score uses a separate aggregation across eligible benchmark evidence.