智能体与自主科学
突破级
暂无讲解视频
收录解读
SkillSafetyBench: Evaluating Agent Safety under Skill-Facing Attack Surfaces 关注的是一个可复用的 AI 系统或评测问题,而不是单点 demo。
Runnable benchmark for safety failures induced by malicious or compromised agent skills.
It targets the skill layer as an attack surface, matching the repository focus on skill systems and agent safety evaluation.
它没有更高,是因为这些新 arXiv 工作仍需要更多独立复现、真实系统部署和长期社区采用来确认影响。