# Cyber Model Arena: Wiz scores 59 AI agents on offensive tasks

> Cyber Model Arena measures whether an AI agent can produce a checkable effect on a real target inside 1,000 seconds.

- **Severity:** Info: Research, tools, evals or lab claims with no direct exposure. Context, not an alert.
- **Category:** Benchmark, Research
- **Published by:** Wiz
- **Disclosed:** 2026-11-10
- **Affects:** Claude Code, OpenAI Codex, Gemini 4 Argon, Claude Opus 5, GPT-6 Sol
- **Research by:** Jay Tenenbaum
- **Primary source:** https://www.wiz.io/blog/cyber-model-arena-ai-benchmark
- **Page:** https://theinjection.dev/items/wiz-cyber-model-arena-benchmark/

Wiz relaunched Cyber Model Arena, a benchmark that runs 59 agent configurations against 327 offensive security challenges with deterministic graders, and reports that the harness around a model shifts its score by about 5 points.

## What happened

Wiz published the relaunched Cyber Model Arena on November 10, 2026, in a post by Jay Tenenbaum that doubles as a guide to building this kind of benchmark. Version 3.9 pairs 26 models with four general purpose harnesses to make 59 agents, then runs each of them three times against 327 challenges.

## How it works

Challenges are drawn from existing real-world sources rather than generated or labelled by a model, and split across code vulnerabilities, API security, web security and cloud security. Deterministic code graders inspect the environment for effects after the agent stops, which avoids using a model as a judge. Agents are isolated from the server and the internet during a run, so they cannot look up answers, and each attempt is capped at 1,000 seconds.

## Who is affected

Teams picking an agent for offensive security work, and defenders estimating what an attacker with an off-the-shelf harness can already do. Wiz reports scores for the ADK React, Claude Code, Antigravity and Codex harnesses paired with Gemini, Claude, GPT, DeepSeek, GLM, Kimi, Grok and MiMo models.

## What to do

Judge a model and its harness together rather than separately: across 22 models run on both Claude Code and ADK React, Wiz measured a mean absolute gap of 5.1 pass@1 points. Wiz also warns that a benchmark with fewer than 30 tasks carries a confidence interval of roughly 18 points at a 50% success rate.

## Average pass@1 by task category, Cyber Model Arena v3.9

Averaged across all 59 agents. Higher means the agent succeeded more often

| Result | Score |
|---|---|
| API security | 78% |
| Web security | 64.6% |
| Cloud security | 56.5% |
| Code vulnerabilities | 38.2% |

Source: https://www.wiz.io/blog/cyber-model-arena-ai-benchmark

## Top agents by pass@1, Cyber Model Arena v3.9

Higher means the agent solved more offensive security challenges

| Result | Score |
|---|---|
| ADK React, Gemini 4 Argon | 77.2% |
| Antigravity, Gemini 4 Argon | 77% |
| ADK React, Gemini 3.8 Flash Cyber | 74.9% |
| Claude Code, Claude Opus 5 | 71.2% |
| Codex, GPT-6 Sol | 71.2% |
| Claude Code, Gemini 4 Argon | 70.8% |
| Claude Code, GPT-6.1 Sol | 70.7% |
| Codex, GPT-6.1 Sol | 70.6% |
| Claude Code, GPT-6 Astra | 70.2% |
| Codex, GPT-5.6 Sol | 69.8% |

Source: https://www.wiz.io/blog/cyber-model-arena-ai-benchmark

## Quick facts

- **Benchmark:** Cyber Model Arena v3.9, 26 models and four harnesses
- **Scale:** 59 agents, 327 challenges, three attempts each
- **Budget:** 1,000 seconds per attempt
- **Grading:** Deterministic code graders, no model used as judge
- **Task mix:** 221 code vulnerability, 54 API, 29 web, 23 cloud
- **Cheapest on frontier:** Codex with GPT-6 Luna, 64.5% at $0.05 per solve
- **Published by:** Jay Tenenbaum, Wiz

## Sources

1. [Wiz Research](https://www.wiz.io/blog/cyber-model-arena-ai-benchmark)
2. [Cyber Model Arena](https://www.wiz.io/cyber-model-arena)