Researchers present a library of attack traces aimed at reusable skills for language-model agents. The work argues that risks in agent skills surface through model decisions, user context, tool calls, and execution feedback rather than through fixed signatures or a single sandboxed run, which limits existing static and dynamic analysis methods. The library is intended to help catalog and study these behaviors.
A new arXiv paper presents GitSkills, a dataset built from GitHub repositories that package agent skills as folders containing a SKILL.md instruction file, sometimes with helper scripts and reference material. The work focuses on skills that language-model agents load when they decide a task matches a skill's description. It is a replacement submission (v2) in the cs.AI category.
A new arXiv preprint describes a method for automatically producing reusable skills that large language model agents can call on to handle complex tasks. The approach uses reinforcement learning to build these procedural units up in stages, aiming to raise the quality of skills derived from documents or past experience. The work is a revised submission (v2) and focuses on skill generation as a modular component of agent design.
A new arXiv paper investigates how language model agents can draw on libraries of reusable knowledge when tackling long-horizon tasks. It contrasts two approaches—subagents and agent skills, where skills are packaged as multi-file bundles—and examines which executes such knowledge more effectively. The study was announced in the cs.AI category and cross-listed in cs.CL and cs.LG.