The authors first compared combinations of publicly available Skills with matched single-Skill controls.
Individually approved agent skills can combine to break workflow policies
Zheng and colleagues show that small additions to an agent’s skill library increase policy violations without changing existing skills or runtime permissions.
Research Lab
Yujia Zheng · Tianhao Li · Shangzhi Xu · Yuekai Huang · Ziyou Jiang · Zhaoyang Li · +3 more
University of Science and Technology Beijing · ZODA · Fangcun Leap · University of New South Wales · Institute of Software, Chinese Academy of Sciences
Research Digest··2 min read
Zheng and colleagues investigate how reusable agent capabilities, called Skills, can become unsafe when combined, even after passing individual security checks.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.