Cybersecurity's AlphaGo moment, or starting from CyberArena's 466 agent matches
网络安全的AlphaGo时刻，或从这466场Agent对局开始
A research team from 2077AI, UC Berkeley RDI and other institutions launched the symmetric attack-defense evaluation platform CyberArena; 12 frontier model and harness combinations went through 22 batches and 466 valid matches. OpenAI's GPT-5.6 Sol ranked first with 78.6 points, Anthropic's Claude Opus 5 second with 73.8 points, and Zhipu's open-source model GLM-5.3 third with 68.5 points. The platform uses an attack-defense capture-the-flag format, with 3 rounds per match; the agent must simultaneously break into the opponent's services and defend the availability of its own business. Comparative experiments showed that on 30 problems that GPT-5.6 Sol, Claude Opus 4.8 and GLM-5.3 could all fully solve in single-player mode, after entering two-way confrontation the attack capture rates in round 3 fell to 52%, 37% and 35% respectively. The paper is arXiv 2610.19005, and the code has been open-sourced on GitHub.
After problems that were almost fully solved in static answering were moved into two-way confrontation, capture rates fell sharply; the 466 matches provide a quantitative reference for the gap between single-machine leaderboards and real-world performance.