---
title: Tip: scan agent skills and MCP servers before you install them; NVIDIA's free SkillSpector flags prompt injection, exfiltration and tool poisoning | ctxsec
url: https://ctxsec.com/items/tip-scan-agent-skills-before-install-skillspector
description: ""
status: "tip; tool is NVIDIA open source (Apache 2.0); figures are NVIDIA's own"
published: 2026-10-27
---

> 

[← Tips & Guides](/tips) · [markdown](/items/tip-scan-agent-skills-before-install-skillspector.md)

tip; tool is NVIDIA open source (Apache 2.0); figures are NVIDIA's own

2026-10-27

### Tip: scan agent skills and MCP servers before you install them; NVIDIA's free SkillSpector flags prompt injection, exfiltration and tool poisoning

Skills and MCP servers run with your coding agent's permissions and usually skip any review. NVIDIA's open-source **SkillSpector** (trending this week) says that in the 31,132-skill subset of its research dataset, 26.1% of skills had vulnerabilities and 5.2% showed likely malicious intent.

- **Before installing:** run `skillspector scan <repo-or-folder> --no-llm` for a fast static pass. It checks 71 patterns across 17 categories, including prompt injection, data exfiltration, MCP tool poisoning, MCP least privilege and memory poisoning. Output is terminal, JSON, Markdown or SARIF.
- **Gate it:** add the SARIF output to CI for any repo that ships skills or `AGENTS.md` files, and use a baseline so re-scans only show new findings.
- **Remember:** a clean scan isn't proof of safety. Still pin skill versions, read anything that asks for shell, network or credential access, and never launch agents from `~/Downloads` (see [the AGENTS.md hijack](/items/forever-agents-md-downloads-folder-agent-hijack)).

Sources: [NVIDIA: SkillSpector (GitHub)](https://github.com/NVIDIA/SkillSpector) · [NVIDIA docs: Scan agent skills before installation](https://docs.nvidia.com/skills/scanning-agent-skills)