“Research in early 2026 shows that approximately 1 in 3 (roughly 26% to 36.8%) of community-developed Claude Skills contain at least one security flaw or risk pattern.”