Agent Skills: one in seventeen requests the agent to dispense with approval, according to the analysis of 1.53 million files

Turning the checks around, acting without asking: about 88,000 officer skills contain this type of instructions, often copied from a handful of original models.

"Colls" are packages of reusable instructions that AI agents load on demand. One line in a highly copied skill can therefore reach all the agents who use it. Timofey Sanko and Yuan Tian have studied for the first time on a large scale the instructions that release human supervision: move forward without asking for agreement, skip controls with the --no-verify option, continue without making any point.

Unconditional and contagious instructions

Of the 1.53 million English skills from GitSkills, about one in seventeen (nearly 88,000) contain such a directive, most often through permission options. 70.1% apply unconditionally, not only in specific contexts such as continuous integration. Skills that execute shell, install packages or use sudo are 1.7 to 1.8 times more likely to contain them, without compensating for additional safeguards.

The phenomenon is propagated by copy: 2.4 per cent of the families of competence involved account for half of the copies. Open source scanners detect at best 46% of these directives, with an accuracy of less than 0.4; the same language model, guided by the authors' annotation grid, rises to 61% recall for an accuracy of 0.71.