CyberArena
Agents · 2026-10-13
CyberArena (2077AI Open Source Foundation, UC Berkeley RDI and collaborators) pits two AI agents, each running in a command-line agent harness, against each other on the same live vulnerable service for three 20-minute rounds: each attacks the opponent's copy to capture rotating flags while patching its own and keeping it passing health probes. A game score is the mean of the agent's capture rate, its defended-flag rate and its service uptime; the published standing is a Bayesian six-axis attack-defense model's posterior expected game score averaged over the other entrants and 40 reference CTF and open-source challenges. Standings are relative to the roster and challenge set.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.6 Sol | 78.6 |
| Claude Opus 5 | 73.8 |
| GLM-5.3 | 68.5 |
| Gemini 3.7 Flash | 67.1 |
| Grok 4.6 | 65.5 |
| Opus 4.8 | 65.0 |
| Gemini 3.5 Flash | 60.6 |
| DeepSeek V4 Pro 0813 | 59.5 |
| MiniMax M3 | 59.1 |
| Kimi K3 | 57.0 |
| Hy4 Preview | 56.3 |