A preprint on arXiv.org titled Beyond Maliciousness: Static and Runtime Attack-Surface Profiling of Agent Skills presents a two-phase attack-surface profiling methodology applied to 6,309 listings from two marketplaces.[1] The authors report that half the corpus carries evidence of shell-execution capability, with 35% showing filesystem read, 31% each showing filesystem write and network egress, and 13% showing credential access.[1] Seven listings in ten carry at least one elevated capability, the authors wrote.[1]

The authors wrote that this is the surface a user accepts at install time, and that almost none of it is visible then.[1] They also found that the permission model is close to inert.[1] Authors may list the tools a skill pre-approves in its own frontmatter, which removes confirmation prompts rather than restricting what the skill can reach.[1] Of the 6,061 listings whose frontmatter parses, 770 declare anything at all, and nothing checks what the declarations express against the bundle.[1]

Declaring is also not declaring accurately, the authors reported: 59% of declaring listings understate what their bundle can do.[1] ClawHub declares roughly twice as often as skillsmp, yet under-declares on 88% of its declarations against 31% for skillsmp.[1] The authors conclude that triage at the marketplace boundary cannot rest on what a listing declares, and must rest on what the bundle can do.[1] They position capability-oriented attack-surface profiling as a necessary complement to maliciousness detection.[1]