Wiz Cyber Model Arena v3.9

Wiz Research's Cyber Model Arena, version 3.9 of its Wiz Offensive AI Benchmark, runs agents, each a model in a general-purpose harness (Wiz's ADK React loop, Claude Code, Codex or Antigravity CLI), on 327 offensive-security challenges in four categories: 221 code-vulnerability tasks (name a known vulnerability's root cause in Python, Java or Go source, or exploit a live 1-day CVE), 54 black-box API security labs, 29 websec.fr web CTFs and 23 cloud misconfiguration CTFs. Agents get only the task instruction, run in isolated Docker containers with no CVE databases or external resources, and are stopped after 1,000 seconds; grading is deterministic (API hooks, exact flags, exact root-cause strings, checks that the vulnerable path ran). Each model, harness and challenge runs three times at high reasoning effort, and the headline is overall pass@1, the equal-weight mean of the four categories' pass@1 rates; scores are for a model in a harness, not for the model alone.