Scan skills with SkillSpector — don’t badge MCP listings “safe”

Scan skills. Handshake MCP. NVIDIA’s SkillSpector is a static linter for agent skill packs (SKILL.md, zips, Git repos). It catches hidden instructions, exfil patterns, CVEs, and tool-poisoning copy. It does not execute the skill, and it does not prove a live HTTPS MCP is safe.

Influzer is pointing builders at the repo — and we are not putting a SAFE badge on directory listings. Most MCP servers have no cloneable tree; many legitimate tools look “excessive” to a skill scanner. Our quality bar stays MCP-shaped: live tools/list (401 counts as alive), indexed tools over stars, Discovery as search-not-install, and a poison pass on submissions (hidden Unicode, HTML comments, instruction-override language) without auto-deleting the catalog.

Full split, the 20-minute rollout, and what a score actually means: Scan agent skills before you install them.