Skip to content

Translation in progress Some entries still include text in the original language.

Failure-First·Indexed by this site · Originally published

ClawKeeper:面向 OpenClaw 智能体的技能、插件与 Watcher 三层安全防护

[Daily Paper] ClawKeeper: Comprehensive Safety Protection for OpenClaw Agents Through Skills, Plugins, and Watchers

AI summary

ClawKeeper 为开源自主智能体运行时 OpenClaw 提出三层实时安全框架,应对敏感数据泄露、权限提升与恶意第三方技能执行。技能层把安全策略注入提示词上下文,插件层在工具调用时检测异常,Watcher 层则是独立于智能体信任边界、可实时干预的系统级中间件。论文称三层防护强于任一单层,Watcher 层对绕过上下文策略的攻击尤其有效,且不明显损害正常任务上的表现。

Read originalfailurefirst.org