LATIDIA · Ciberseguridad

Cuando las habilidades les indican a los agentes que no pregunten: Un estudio a gran escala de las directivas de supervisión en las habilidades de los agentes

arXiv:2610.37734v1 Anuncio Tipo: nuevo Resumen: Las habilidades de los agentes, paquetes de instrucciones reutilizables centrados en un archivo SKILL.md, extienden los agentes de IA bajo demanda, y una sola línea en una habilidad ampliamente copiada llega a cada agente que carga

La noticia

arXiv:2610.37734v1 Announce Type: new Abstract: Agent skills, reusable instruction packages centred on a SKILL.md file, extend AI agents on demand, and a single line in a widely copied skill reaches every agent that loads it. Some lines are oversight-loosening directives: they tell the agent to proceed without approval, to bypass checks with --no-verify, or to keep working without checking in. Although such directives can remove a checkpoint from every agent that loads them, no prior work has examined them. We present the first large-scale study of oversight-loosening directives, analyzing 1.53 million English skills from GitSkills. We split skills into units (sentences, bullets, table cells, and code lines), develop a codebook