# CyberBench v1.1 Contamination Audit

> CyberBench v1.1 scores 7.5/10 (Tier A, Low Risk) on BenchAudit as of November 2026, reflecting a strictly private held-out test set and isolated offline third-party grading.

- **Canonical URL:** https://benchaudit.org/benchmarks/cyberbench/
- **Benchmark ID:** `cyberbench`
- **Hub:** Vals AI
- **Category:** Coding
- **Domains:** Cyber, Coding
- **Skills:** Agentic, Coding, Tool Use, Reasoning
- **BenchAudit Score:** 7.5/10 (Tier A - Low Risk)
- **Last Audited:** 2026-11-22 (Ref 2026-11)

## Factor Scores

| Factor | Score (/10) | Weight |
| --- | ---: | ---: |
| Privacy | 10 | 40% |
| Originality | 0 | 25% |
| Freshness | 10 | 20% |
| Grading | 10 | 15% |

## Primary-Source Evidence

### Privacy
> "This dataset remains private at all times, and is the only dataset used for the proprietary benchmark results we publish. It is private to prevent leakage into the training data of foundation models."
- Source: https://www.vals.ai/methodology

### Originality
> "Agents are tested on OSS-Fuzz regressions from ARVO images. In cybersecurity, fuzzing is an automated testing technique that feeds many inputs into a program to trigger crashes and memory-safety bugs."
- Source: https://www.vals.ai/benchmarks/cyber

### Freshness
> "CyberBench v1.1 supersedes v1 because the results were rerun with offline task sandboxes and a hardened clean-room grader. It keeps the 60-task PoC track and publishes a 56-task Patch roster containing the tasks that can be graded reliably under that setup."
- Source: https://www.vals.ai/benchmarks/cyber

### Grading
> "CyberBench v1.1 supersedes v1 because the results were rerun with offline task sandboxes and a hardened clean-room grader. It keeps the 60-task PoC track and publishes a 56-task Patch roster containing the tasks that can be graded reliably under that setup."
- Source: https://www.vals.ai/benchmarks/cyber

## Top Frontier Models on CyberBench v1.1

| Rank | Model | Creator | Percentile |
| ---: | --- | --- | ---: |
| #1 | [Muse Wonder](https://benchaudit.org/models/muse-wonder__max/) | Meta | 89.6p |
| #2 | [GPT-6 Sol](https://benchaudit.org/models/gpt-6-sol__max/) | OpenAI | 86.6p |
| #3 | [Gemini 4 Argon](https://benchaudit.org/models/gemini-4-argon__high/) | Google | 83.7p |
| #4 | [GPT-6 Luna](https://benchaudit.org/models/gpt-6-luna__max/) | OpenAI | 80.7p |
| #5 | [Claude Haiku 5.5](https://benchaudit.org/models/claude-haiku-5-5__max/) | Anthropic | 77.7p |
| #6 | [MiMo-V2.6-Flash](https://benchaudit.org/models/mimo-v2-6-flash/) | Xiaomi | 74.7p |
| #7 | [DeepSeek V4.1 Flash 1017](https://benchaudit.org/models/deepseek-v4-1-flash-1017__max/) | DeepSeek | 71.7p |
| #8 | [Muse Wonder Full](https://benchaudit.org/models/muse-wonder-full__max/) | Meta | 68.7p |
| #9 | [MiMo-V2.6-Pro](https://benchaudit.org/models/mimo-v2-6-pro/) | Xiaomi | 64.3p |
| #11 | [Inkling 2 Small](https://benchaudit.org/models/inkling-2-small/) | Thinking Machines | 58.3p |
