SkillSpector is a security scanner NVIDIA built to answer one question before an AI coding agent installs a third-party skill: is this safe to run? Skills for Claude Code, Codex CLI, Gemini CLI, and similar harnesses execute with implicit trust and little vetting, and the README cites research finding 26.1% of skills in a 42,447-skill sample contain at least one vulnerability and 5.2% show likely malicious intent. SkillSpector runs a two-stage pipeline: a fast static-analysis pass combining regex pattern matching across 11 analyzers, Python AST inspection for dangerous calls (exec, eval, subprocess, dynamic imports), YARA signature matching, and taint tracking that follows data from a source (env vars, file reads, network input) to a sink (network output, exec/eval); then an optional LLM semantic pass that evaluates context and intent to filter false positives, which the README states raises precision to roughly 87%. Detection covers 71 named patterns across 17 categories -- prompt injection, anti-refusal jailbreak framing, data exfiltration, privilege escalation, supply-chain issues (unpinned or typosquatted dependencies, curl-pipe-bash, live CVE lookups via OSV.dev), excessive agency, unsanitized output handling, memory poisoning, tool misuse, self-modifying rogue-agent behavior, and MCP-specific least-privilege and tool-poisoning checks such as hidden Unicode instructions in tool metadata. It scans a local directory, a SKILL.md file, a zip file, or a remote Git URL, exporting terminal, JSON, Markdown, or SARIF reports; a baseline/suppression mechanism lets a team accept known findings by glob rule or exact fingerprint so re-scans only surface genuinely new issues. It also runs as an MCP server so an agent can call scan_skill as a tool and gate installs on the returned risk score at runtime, rather than as a separate audit step -- though the project is explicit that it never executes the scanned skill itself. LLM analysis is pluggable across OpenAI, Anthropic, AWS Bedrock, NVIDIA's build.nvidia.com, or a locally authenticated Claude/Codex CLI session, with a --no-llm flag to keep everything static and local. The scanner is part of NVIDIA's broader Verified Skills pipeline that vets and signs skills before they reach the NVIDIA skills catalog.