Skip to main content
BenchLM
Data

CyberGym

We show this table for reference; we do not rank on it.

Data verified

A cybersecurity task benchmark for evaluating defensive cyber workflows and vulnerability-oriented agent performance.

Benchmark score on CyberGym — November 19, 2026

We compile the CyberGym rows from benchmark-owner or independent runs, provider self-reports and secondary reports. MiMo-V2.6-Flash leads the table at 95.1%, as reported by Xiaomi, followed by MiMo-V2.6-Pro (94.0%, as reported by Xiaomi) and DeepSeek V4.1 Flash 1017 (88.1%, as reported by DeepSeek). We do not use these results to rank models overall.

29 modelsAgenticCurrentDisplay onlyUpdated November 19, 2026

Benchmark score table (29 models)

Of 29 results: 22 lab-reported, 7 benchmark-run.

Ranks on this table count the 7 benchmark-run and independent results. A lab-reported result keeps its place by score, without a rank.

Lab-reported
22
Independent
0
Benchmark-run
7
Who ran it

Showing all 29 results, in score order.

Benchmark score results for CyberGym
RankModel / configurationScoreParameters (B)Open / closed
No rank on this tableMiMo-V2.6-FlashXiaomiReported by Xiaomi
95.1%
Not reportedOpen
No rank on this tableMiMo-V2.6-ProXiaomiReported by Xiaomi
94.0%
Not reportedOpen
No rank on this tableDeepSeek V4.1 Flash 1017DeepSeekReported by DeepSeek
88.1%
Not reportedOpen
No rank on this tableLing 3.1 FlashInclusionAIReported by InclusionAI
87.9%
Not reportedOpen
No rank on this tableFugu CyberSakana AIReported by Sakana AI
86.9%
Not reportedClosed
No rank on this tableAtria Dawn PreviewShanghai Artificial Intelligence LaboratoryReported by the lab
86.5%
Not reportedOpen
No rank on this tableGemini 3.8 Flash CyberGoogleReported by Google
86.2%
Not reportedClosed
No rank on this tableStep 5 PreviewStepFunReported by StepFun
84.7%
Not reportedOpen
No rank on this tableGPT-5.6 SolOpenAIReported by OpenAI
84.5%
Not reportedClosed
No rank on this tableGLM-5.3Z.AIReported by Z.AI
84.5%
Not reportedOpen
No rank on this tableClaude Mythos 5AnthropicReported by Anthropic
83.8%
Not reportedClosed
No rank on this tableDeepSeek V4 ProDeepSeekReported by DeepSeek
83.3%
Not reportedOpen
No rank on this tableGemini 3.5 Flash CyberGoogleReported by Google
83.2%
Not reportedClosed
1Claude Mythos PreviewAnthropicRun by CyberGym
83.1%
Not reportedClosed
No rank on this tableMiniMax M3.1 FlashMiniMaxReported by MiniMax
82.2%
Not reportedClosed
No rank on this tableGPT-5.6 TerraOpenAIReported by OpenAI
81.8%
Not reportedClosed
No rank on this tableGPT-5.5OpenAIReported by OpenAI
81.8%
Not reportedClosed
No rank on this tableGPT-5.4OpenAIReported by OpenAI
79.0%
Not reportedClosed
No rank on this tableHy4 previewTencentReported by Tencent
78.4%
Not reportedOpen
No rank on this tableGPT-5.6 LunaOpenAIReported by OpenAI
77.9%
Not reportedClosed
No rank on this tableDeepSeek V4 Flash 0731DeepSeekReported by DeepSeek
76.7%
Not reportedOpen
No rank on this tableClaude Opus 4.7AnthropicReported by OpenAI (comparison table)
73.1%
Not reportedClosed
2GLM-5.1Z.AIRun by CyberGym
68.7%
Not reportedOpen
3Claude Opus 4.6AnthropicRun by CyberGym
66.6%
Not reportedClosed
4Claude Sonnet 4.6AnthropicRun by CyberGym
65.2%
Not reportedClosed
No rank on this tableMuse Spark 1.1MetaReported by Meta
59.0%
Not reportedClosed
5Claude Opus 4.5AnthropicRun by CyberGym
50.6%
Not reportedClosed
6Muse SparkMetaRun by CyberGym
43.5%
Not reportedClosed
7GLM-5Z.AIRun by CyberGym
43.2%
Not reportedOpen

A dash in the rank column marks a lab-reported result. It keeps its place by score, without a rank on this table.

  • Lab-reported
  • Benchmark-run
  • Independent
What the labels mean

Results reported by model developersUnranked

Labs ran these themselves, with a metric, split or setting the table above does not use. They take no rank here.

CyberGym results reported by model developers, unranked
MAI-Cyber-1-FlashMicrosoft90.4%Target any-ofCyberGym Level 1MDASH harness with GPT-5.4 as the escalation model, a two-model systemNot comparable to the scores aboveReported by Microsoft

A lab’s own result is shown for reference. It never takes a rank on this board.

How we label sources

Among the reported CyberGym rows, MiMo-V2.6-Flash is first at 95.1%, as reported by Xiaomi. The third row is 7.0 points behind. The broader top-10 range is 10.6 points, so the table still separates the published systems.

29 models have been evaluated on CyberGym. The benchmark falls in the Agentic category. CyberGym is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.

About CyberGym

Year

2026

Tasks

1,507 vulnerability analysis instances

Format

Vulnerability reproduction and PoC generation

Difficulty

Real-world cybersecurity

CyberGym includes 1,507 benchmark instances from historical vulnerabilities across 188 large software projects. BenchLM stores CyberGym as a display-only agentic security benchmark when exact provider comparison values are published.

CyberGym is one of the cyber benchmarks we track. The cybersecurity benchmark comparison sets it beside the others, one table per security job.

Freshness and provenance

Version

CyberGym 2026

Refresh cadence

Quarterly

Staleness state

Current

Question availability

Public benchmark set

CurrentDisplay only

BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.

Questions

What does CyberGym measure?

A cybersecurity task benchmark for evaluating defensive cyber workflows and vulnerability-oriented agent performance.

Which model scores highest on CyberGym?

MiMo-V2.6-Flash by Xiaomi currently leads with a score of 95.1% on CyberGym, as reported by Xiaomi.

How many models are evaluated on CyberGym?

29 AI models have published results on CyberGym in the BenchLM catalog.

Last updated: November 19, 2026 · BenchLM version CyberGym 2026

Know when it’s worth switching models

The model to choose, the cheaper alternative, and the release we would wait on.

Read a sample issue

Join 7,000+ subscribers.

One email each week. Unsubscribe anytime.