KKman scans unfamiliar agent skills with NVIDIA's open-source SkillSpector before installing; Anthropic's own docx skill scored 86, and reading the flagged lines showed normal file handling
A skill is an instruction file for your AI assistant, so installing a stranger's skill means letting a stranger give your assistant orders. KKman (@JackLuiKK) cloned NVIDIA's open-source scanner SkillSpector (Apache-2.0) on Linux and ran static checks only on three samples: the repo's clean example scored 0, its malicious example 100, and Anthropic's official docx skill 86 with a DO NOT INSTALL verdict. Opening the flagged lines, he found normal steps for repacking Word files and launching LibreOffice. His takeaway: the score is where you start, and you still read the flagged lines. We reran the same three scans and got the same scores, issue counts and coverage, then scanned two newer skills and saw the same pattern.
Visit site SkillSpector (NVIDIA) ↗- Pricing
- Free
- Region
- Global
- Where it's from
- Chinese community
- Added
- Nov 14, 2026
- Tested on
- Nov 14, 2026
- Version we used
- NVIDIA/SkillSpector main ce46d46 (CLI v2.12.0), static mode (--no-llm); author used 17c4d38
- Categories
- CodingAutomation & Agents
- Original author
- KKman @JackLuiKK ↗
- Our test
- Tested by us
- Cost / time
- SkillSpector is free and open source. Static checks run locally at no model cost; the optional model comparison is billed by whichever provider you configure.
The task
Before installing someone else's skill from GitHub or a marketplace, you want to know whether it hides instructions that change your assistant's behavior, read environment variables or send data out.
Steps
- Install Python 3.12+ and uv, then run uv tool install git+https://github.com/NVIDIA/skillspector.git, or clone the repo and install it in a virtualenv.
- Download the skill and scan only the folder you will actually install: skillspector scan ./skill-folder --no-llm. --no-llm means static checks only, no model calls.
- To keep a report, add --format json --output report.json and look at risk_assessment: a 0–100 score, a severity and a recommendation (SAFE / CAUTION / DO_NOT_INSTALL).
- Open the file and line for each finding and decide what it does. In the docx skill, the "tool misuse" finding came from a step that deletes the old output file before zipping, and "environment variable harvesting" from copying the environment before launching LibreOffice.
- If you want it to compare what the skill says it does with what its code does, configure a model provider (OpenAI, Anthropic, local Ollama and others) and rescan without --no-llm.
Prompt
skillspector scan ./skill-folder --no-llm --format json --output report.json
What the author says
KKman @JackLuiKK (X post, 13 Nov 2026): scanned SkillSpector main at commit 17c4d38 (CLI shows v2.12.0) with static checks only, without running any script in the scanned skills. The safe-greeting example took about 2.2 s: 0/100, SAFE. The malicious example chef-assistant took about 2.0 s: 100/100, DO NOT INSTALL, 7 findings (prompt injection, environment variable harvesting, sending data out). Anthropic's official docx skill (anthropics/skills commit 2c5958e) took about 27.5 s: 61 files, 91.8% coverage, 86/100, CRITICAL, DO NOT INSTALL, 11 findings including 3 high. After reading the code he judged them normal document-conversion steps, not covert exfiltration. The paper he cites, Agent Skills in the Wild, analyzed 31,132 skills and found 26.1% had at least one vulnerability pattern; that is the paper's number, not this tool's.
Our test
On 14 Nov 2026 on Linux we cloned NVIDIA/SkillSpector (main ce46d46, CLI shows v2.12.0), installed it in a uv virtualenv on Python 3.13.5, and used --no-llm static checks throughout. The repo's safe_skill (safe-greeting): 0, LOW, SAFE, 0 findings. malicious_skill (chef-assistant): 100, CRITICAL, DO_NOT_INSTALL, 7 findings. docx from anthropics/skills commit 2c5958e: 86, CRITICAL, DO_NOT_INSTALL, 11 findings, 61 components, 91.8% coverage, matching the author. We also scanned two more. op7418/guizang-game-skill (6001e56): 56, HIGH, DO_NOT_INSTALL, 6 findings; the 3 high ones all say a referenced script could not be fully parsed (for example build_single_html.mjs), and the report itself notes that incomplete analysis is not proof of malice; 2 low ones come from the LICENSE text. globalmpc/backendskills, skills folder only: 0, no findings, but 66.7% coverage and a CAUTION recommendation; the whole repo including benchmark logs: 100 with 145 findings.
What we found when we tried it:The author's three results reproduce exactly. The scanner catches the obvious malicious sample, but often gives high scores to normal skills too, mostly for calling external commands or scripts it couldn't fully parse. Use it the way the author does: scan only the folder you will install, then read each flagged line.
How to get in
The original post is in Chinese on X (use browser translate). The tool, its docs (https://docs.nvidia.com/skills/scanning-agent-skills) and its output are in English.
Who it suits
Claude Code, Codex and Gemini CLI users who often install skills from GitHub or marketplaces, especially skills that ship scripts.
Pitfalls
- A high score isn't proof of malice: the official docx skill scored 86 and the game skill we scanned 56, and the flagged lines were normal steps or scripts the scanner couldn't fully parse.
- A low score isn't proof of safety: with static checks only, risks that need reading intent can slip through, as SkillSpector's own docs say.
- Scanning a whole repo counts test data and logs too: the full backendskills repo got 145 findings and 100/100, while its skills folder alone got 0.
- The CLI version (v2.12.0) doesn't match the latest release (v2.13.x); both the author and we saw this.
More in this category
- A demo video per docs page that flags itself when the UI changesVideoScreenplayClaude CodeTool free; uses your agent's quota◐ Partly testedOverseas
- AI backend code that guards against races, leaks and bad migrationsCodingbackendskillsClaude CodePostgreSQLFree; about 2× tokens per task◐ Partly testedOverseas
- An afternoon of AI work, a 3D action game you play in the browserCodingClaude CodeTripoJimengSkill free; Claude quota plus asset tools◐ Partly testedChina
- Keep a long-running AI-built project from turning into a messCodingClaudeCodexoil-delegateNeeds Claude and Codex plans◐ Partly testedChina
- Use a web app once and get an AI write-up of how it worksResearchREAClaude CodeChromeFree tools; needs Claude or Codex◐ Partly testedOverseas