CyberArena：在实时对称攻防对抗中评测 AI 智能体
原标题：CyberArena: Benchmarking AI Agents in Live Symmetric Attack-and-Defense Combat
AI 摘要
UC Berkeley RDI 等机构的研究者提出 CyberArena，一个面向 AI 安全智能体的实时对称攻防基准。与传统的单角色、固定目标 jeopardy 式测试不同，CyberArena 让两个智能体在三个回合中互相攻击对方存在漏洞的服务，同时修补自己的服务，并由竞技场独立验证每次攻击和健康检查。研究者在 12 个前沿模型、40 个挑战上进行了 466 场比赛，结果显示这种对抗形式能区分出单角色测试无法区分的模型能力。
正文节选
CyberArena Leaderboard Mirror of the CyberArena attack-defense CTF leaderboard By Bangya Liu, Yueyang Pan and Hanhong Zhao · 2077AI · UC Berkeley RDI 📄 Paper | 🏆 Live leaderboard | 💻 Code (Octagon) | 🧩 Challenges | ▶️ Match replays | 📚 alphaXiv Most cybersecurity benchmarks for AI agents are exams: one agent, one fixed target, one role. CyberArena replaces the exam with a contest. Two agents receive identical copies of a vulnerable service and, over three rounds, each attacks the other's copy wh