Vibe Coding
Pattern: Skill Supply-Chain Defense Is Moving to the Client Boundary, Not the Registry
Claude Code 2.1.228 sanitizes and labels skill descriptions synced from claude.ai and strips their ability to run `!` commands or expand `@` files locally — enforcement at the point of execution, on the user's machine. That pairs with Anthropic's enterprise skill-and-plugin security scanning (beta, 08-06) at upload time, giving two independent layers rather than trusting a marketplace review. The design lesson for anyone shipping a skill distribution channel: assume the registry will be poisoned and put the real check where the skill actually runs.
↳ Follow the thread