Vercel Skills 達 62,000 個並啟安全審核

🔑 增強重點摘要

• •Vercel Skills ecosystem reached 62,000 skills and 2 million npx CLI installs, with top skills from Vercel achieving a combined 66,000 installs[5].
• •Skills.sh serves as an open directory for reusable AI agent skills, installable via CLI into agents like Cursor and Claude, with rapid adoption post-launch reaching tens of thousands of installs[8][6].
• •Security enhancements include partnerships with Socket and Snyk for audits, plus a new integration with Gen's Agent Trust Hub (announced Feb 17, 2026) providing four-tier risk ratings (Safe, Low Risk, High Risk, Critical Risk) directly in skills.sh for over 6 million developers[3][4].
• •New Audits leaderboard displays per-skill security assessments to promote transparency and safety[article].
• •Vercel evals show static AGENTS.md (8KB compressed docs) outperforming skills (100% vs 79% pass rate on Next.js tasks), highlighting complementary uses: static for broad knowledge, skills for specific actions[1][7].

🛠️ 技術深入

• •Skills are modular capabilities published to skills.sh, installed via npx CLI (e.g., one-command into 10+ agents like Cursor, Claude), enabling on-demand actions such as commerce (Universal Commerce Protocol) and workflows (durable async functions with retry logic)[6][article].
• •Security audits target threats like remote shells; Gen's ATH uses AI for independent verification classifying skills as Safe (best practices), Low Risk (minor indicators), High Risk (significant concerns), or Critical Risk (severe/malicious)[3][4].
• •Eval data: Skills with default behavior matched baseline 53% pass rate (agent invocation failure in 56% cases); explicit instructions boosted to 79%, but static 8KB AGENTS.md achieved 100% on Next.js 16 API tasks (perfect Build/Lint/Test scores)[1][7].
• •Failure patterns: Agent non-invocation, fragile triggers sensitive to phrasing; skills suit vertical workflows, AGENTS.md for horizontal Next.js knowledge[1].

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線