---
title: Wiz Cyber Model Arena update: best agent solves 77% of offensive security tasks first try, and the harness matters as much as the model | ctxsec
url: https://ctxsec.com/items/wiz-cyber-model-arena-2026-10
description: ""
status: "vendor research (benchmark)"
published: 2026-10-14
---

> 

[← AI](/ai) · [markdown](/items/wiz-cyber-model-arena-2026-10.md)

vendor research (benchmark)

2026-10-14

### Wiz Cyber Model Arena update: best agent solves 77% of offensive security tasks first try, and the harness matters as much as the model

Wiz updated its Cyber Model Arena (Oct 14), which tests model and harness pairings on 327 real-world offensive tasks: 221 code-vuln/1-day CVE challenges, 54 live API security labs, 29 websec.fr CTFs and 23 cloud misconfiguration CTFs. Agents run in isolated Docker containers with no CVE databases. **Gemini 4 Argon** on a plain ADK ReAct loop leads at 77.2% pass@1 (about 4.3 min and $0.77 per solve). Claude Opus 5 + Claude Code and GPT-6 Sol + Codex tie at 71.2%. GPT-6 Luna + Codex reaches 64.5% at about $0.05 per solve. Codex and Antigravity harnesses were newly added. The findings matter for defenders: the same model swings widely by harness (Claude Sonnet 5: 59.3% in Claude Code vs 42% in ADK ReAct), successful attacks typically cost a few dollars and take minutes, and finding and exploiting code vulns remains the hardest category (top score 65.8%). Treat it as one vendor's benchmark, not a field measurement.

Sources: [Wiz Cyber Model Arena](https://www.wiz.io/cyber-model-arena) · [@wiz_io on X](https://x.com/wiz_io/status/2110352227806409035)