安全★★★arXiv · 2026-07-16
Cost-Aware Evaluation of Offensive and Defensive Security Agents
Traditional security agent evaluations focus on peak offensive capabilities with generous resources but overlook operational costs. This paper assesses language-model security agents using a cost-success framework on offensive Cybench and defensive Splunk BOTS v1 challenges.
📌 Key points
- Traditional evaluations emphasize high-budget offensive capabilities
- Each reasoning step incurs operational costs
- Models are compared at fixed cost in this study
本页为 gitzw.com 基于公开来源的 AI 中文解读，非原文转载。