SkillGate Screens Agent Skill Files Before Install: F1=0.817 at 1.13% FPR While Cutting LLM Input Tokens 77%

arXiv 2607.25619·high signalXBlueskyLinkedInCopy link

A paper posted 2026-07-28 (arXiv 2607.25619) presents SkillGate, a security gateway that screens Markdown skill packages before Cursor, Claude Code, or GitHub Copilot install them, using a regex prefilter that lets safe files bypass the LLM judge entirely and sends only matched snippet windows for flagged files. On SkillsBench (n=1,650, 9.1% malicious) it hits F1=0.817 and 1.13% false-positive rate while reducing LLM input tokens 77% versus full-file screening, and beats two existing tools by 5-6x on AUPRC (0.830 vs 0.144 and 0.162). The threat is independently corroborated: Snyk's February 2026 audit of 3,984 ClawHub and skills.sh skills found 534 (13.4%) with a critical issue and 76 confirmed malicious payloads, 8 of which were still live at publication.