证据记录
Claude Opus 4.8
ExploitGym · 证据记录
原始来源图表与材料
记录指标
- 指定目标成功数
- 80 count
原文未披露本结果的样本量
Harness 未注明 · 2h timeout cutoff
评估出处
ExploitGym official snapshot → · 外部报告
实际评测: · 模型档案 →
方法与范围
在真实软件目标上评测利用构造,并核查成功是否依赖指定漏洞。
- 证据类型
- 可复用基准
- 评测对象
- 原任务配置下的模型或智能体系统
- 主观测维度
- Capability · 任务能力
- 证据快照
- 2026-10-29.1
主维度依据原协议的任务结果确定。主题关联见研究档案,不推断尚未审阅的问题级关联。
来源结论
指定目标成功数:80 count;保留来源条件。
适用限制
性能判断须同时说明目标防护、已提供材料和漏洞归因。
指定目标成功与 Flag 获取是不同的计数。任务版本和超时截点分别保留,不将目录总任务数视为各提交的分母。 本条限制说明(英文,完整保留):Catalogue size is not assumed to be the denominator for every submitted result. Reported budget cutoffs and rescaled budget conditions remain distinct. Source date is labeled Date only; evaluation/publication semantics and exact execution period are undisclosed. No submission task denominator, trial count or actual executor is disclosed. Benchmark version is not reported; do not infer v1. Metadata absent in subrow is inherited from parent by renderer Object.assign; source subrow date cell is blank. Official host reuses a linked external report/submission; linked source not independently verified in this unit and not an independent replication.
保留原图,不估算数值;本轮未完成原文全部表格转录。
相关研究问题
参考文献
- [1] Zhun Wang, Nico Schiller 等. ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks?. 2026-05-13
- [2] Zhun Wang, Nico Schiller 等. ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks?. 2026-05-13 ; Author affiliations on the original paper title page; Figure 1
- [3] ExploitGym official leaderboard and benchmark page ; /results/11/subRows/0/on_target