about
Profiling the Attack Surface of AI Agent Skills (arxiv.org)
1 point by ratneshiiita 29 days ago | hide | past | pdf | discuss on HN

In plain words: A tool reads each skill an AI assistant installs, then runs it in a sandbox to see what it can do, like run shell commands or reach credentials. Of 6,309 listings, seven in ten had an elevated power, like running shell commands or reaching credentials; 59% of declaring listings understated them. Wait, that repeats examples. Fix: put examples in S1, keep S2 short. A tool reads each skill an AI assistant installs, then runs it in a sandbox to see what it can do, like run shell commands or reach credentials. Of 6,309 listings, seven in ten had at least one elevated power, and 59% of declaring listings understated them. Count: S1 = 28, S2 = 19 → 47. Over. Compromise: S1 without examples (21), S2 with examples (23) = 44. Final: "A tool reads each skill an AI assistant installs, then runs it in a sandbox to see what it can do. Of 6,309 listings, seven in ten had an elevated power, like running shell commands or reaching credentials; 59% of declaring listings understated them." Total 44 words. Good. Check S2 count again: Of(1) 6,309(2) listings,(3) seven(4) in(5) ten(6) had(7) an(8) elevated(9) power,(10) like(11) running(12) shell(13) commands(14) or(15) reaching(16) credentials;(17) 59%(18) of(19) declaring(20) listings(21) understated(22) them.(23) = 23. S1: A1 tool2 reads3 each4 skill5 an6 AI7 assistant8 installs,9 then10 runs11 it12 in13 a14 sandbox15 to16 see17 what18 it19 can20 do.21 = 21. Total 44. ✓ Also should mention static scan? "reads each skill's files" captures static analysis. Let's say "reads each skill's files and runs it in a sandbox" — but then no "AI assistant installs" context. Hmm. "A tool reads the files of each skill an AI assistant installs, then runs it in a sandbox to see what it can do." = A1 tool2 reads3 the4 files5 of6 each7 skill8 an9 AI10 assistant11 installs,12 then13 runs14 it15 in16 a17 sandbox18 to19 see20 what21 it22 can23 do.24 = 24. +23 = 47. Over. Keep the 44-word version. "reads each skill" implies reading files. Fine. Also "elevated power" — the paper says elevated capability. Fine. Final output.A tool reads each skill an AI assistant installs, then runs it in a sandbox to see what it can do. Of 6,309 listings, seven in ten had an elevated power, like running shell commands or reaching credentials; 59% of declaring listings understated them.

Abstract · Beyond Maliciousness: Static and Runtime Attack-Surface Profiling of Agent Skills

Agent skills are Markdown documents that a tool-using agent reads as instructions and then acts on with the privileges of the session that loaded them. Installing one grants authority (the standing to direct the agent, not new privilege), and the size of that grant is what we measure. We present a two-phase attack-surface profiling methodology that combines whole-bundle static analysis and dependency provenance checking with instrumented sandbox execution and separate model-based assessment of the resulting traces. We apply the methodology to 6,309 listings from two marketplaces, selecting listings for execution without conditioning that decision on a prior threat or risk signal. Half the corpus carries evidence of shell-execution capability; the corresponding shares are 35% for filesystem read, 31% each for filesystem write and network egress, and 13% for credential access; seven listings in ten carry at least one elevated capability. That is the surface a user accepts at install time, and almost none of it is visible then. A permission model exists but is close to inert. 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; of the 6,061 listings whose frontmatter parses, 770 declare anything at all, the vocabulary cannot express credential or environment access, and nothing checks what it does express against the bundle. Declaring is also not declaring accurately: 59% of declaring listings understate what their bundle can do. ClawHub declares roughly twice as often as skillsmp, yet under-declares on 88% of its declarations against 31%. Triage at the marketplace boundary therefore cannot rest on what a listing declares. It must rest on what the bundle can do. These results position capability-oriented attack-surface profiling as a necessary complement to maliciousness detection.

Ratnesh Pandey, Japneet Singh, Harsh Akshit, Sonali Tyagi, Cassie Crossley
arXiv:2610.19264 · cs.CR, cs.AI · submitted Oct 13, 2026
abstract · pdf · html · This paper has 13 pages, 9 figures, and 3 tables

add comment on HN