Parth A. Gandhi and colleagues submitted Scan Before You Install: Detecting and Mitigating Local and Composite Vulnerabilities in Skills to arXiv on 27 October 2026. The authors report a six-module scanner for unintended vulnerabilities within agent skills and vulnerabilities arising when skills interact. This is a research report; the saved abstract identifies no fixed software release.[1]

A skill packages natural-language instructions, scripts, and reference files for an agent based on a large language model. Under user-authorized permissions, skills may access files, the shell, and the network. The authors say existing scanners focus primarily on malicious skills and examine skills individually, missing weaknesses in benign packages and interactions through shared resources, invocations, or instructions.[1]

The authors report that three analysts confirmed 53 of 59 findings across 23 repositories in a survey of 1,587 skills from the 75 most-starred GitHub repositories. They also report two composite findings validated through proof-of-concept attacks. On SCR-Bench, their composite analysis achieved an F1 score of 0.991, compared with 0.816–0.858 for three SkillsGuard configurations. These are author-reported results.[1]

The mitigation module recommends skill substitution, patches, and configuration changes. For 47 of 55 GitHub findings, the authors say mitigations blocked proof-of-concept attacks in all trials while preserving the legitimate functionality they tested. That statement is bounded by those attacks and functionality checks.[1]

[1]: Scan Before You Install: Detecting and Mitigating Local and Composite Vulnerabilities in Skills, arXiv, 27 October 2026.

Sources

  1. Scan Before You Install: Detecting and Mitigating Local and Composite Vulnerabilities in SkillsarXiv · 27 Oct 2026

Check or correct this page

Verification record

The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.

ddd73ca7a0559190bf550e3cfa18e45d64d4b280a0ed5b809c1519b35e4bc652