Atlas
Benchmarks
331 sources · 1,258 models × 2,980 benchmarks · 155,643 scoresWiz Cyber Model Arena v3.9 (Code Vulns)
The Code Vulns category of Wiz's Cyber Model Arena v3.9: 221 challenges in which an agent either identifies the root cause of a known vulnerability pattern in stripped open-source Python, Java or Go code, graded by an exact root-cause string, or exploits a live 1-day (known-CVE) target. Pass@1 averages three attempts per challenge, for a model in a general-purpose harness.
Models
24
Top score
65.8
Leader
Gemini 4 Argon
Updated
—
Resolution
—
Saturation
—
No signal or difficulty estimate. Held out of the fit by design.
Organizations
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Gemini 4 Argon | high · Wiz Cyber Model Arena ADK React; simple ReAct loop built with ADK; single bash tool; no cyber-specific tools; general system prompt | 65.8 | |
| 2 | GPT-5.6 Sol | high · Codex | 57.2 | |
| 3 | Gemini 3.8 Flash Cyber | high · Wiz Cyber Model Arena ADK React; simple ReAct loop built with ADK; single bash tool; no cyber-specific tools; general system prompt | 57.0 | |
| 4 | GPT-6.1 Sol | high · Claude Code | 56.6 | |
| 5 | GPT-6 Astra | high · Claude Code | 56.6 | |
| 6 | Claude Opus 5 | high · Claude Code | 55.2 | |
| 7 | GPT-6 Sol | high · Codex | 55.1 | |
| 8 | GPT-5.6 Terra | high · Codex | 48.9 | |
| 9 | GPT-6 Luna | high · Codex | 46.8 | |
| 10 | Claude Haiku 5.5 | high · Claude Code | 46.6 | |
| 11 | Gemini 3.7 Flash | high · Antigravity CLI | 45.6 | |
| 12 | Opus 4.8 | high · Claude Code | 44.0 | |
| 13 | Gemini 3.6 Flash | high · Claude Code | 39.4 | |
| 14 | DeepSeek V4 Pro 0813 | high · Claude Code | 37.9 | |
| 15 | Gemini 3.5 Flash | high · Claude Code | 37.7 | |
| 16 | DeepSeek V4.1 Flash 1017 | high · Claude Code | 37.1 | |
| 17 | Sonnet 5 | high · Claude Code | 36.3 | |
| 18 | GLM-5.3 | high · Claude Code | 32.9 | |
| 19 | Sonnet 4.5 | high · Claude Code | 32.0 | |
| 20 | Kimi K3 | high · Wiz Cyber Model Arena ADK React; simple ReAct loop built with ADK; single bash tool; no cyber-specific tools; general system prompt | 27.5 | |
| 21 | GLM-5.3 Flash | high · Claude Code | 24.1 | |
| 22 | Grok 4.6 | high · Wiz Cyber Model Arena ADK React; simple ReAct loop built with ADK; single bash tool; no cyber-specific tools; general system prompt | 23.4 | |
| 23 | DeepSeek V4.1 Flash | high · Claude Code | 21.0 | |
| 24 | MiMo-V2.6-Pro | high · Wiz Cyber Model Arena ADK React; simple ReAct loop built with ADK; single bash tool; no cyber-specific tools; general system prompt | 15.2 |
