SkillSpector is a free, open-source scanner from NVIDIA. It reads an AI skill’s files before you install it and gives a risk score from 0 to 100, without running the skill. A low score is not proof of safety, and its default setting sends the files to an AI provider.
- What it is: a command-line scanner for AI agent skills, made by NVIDIA and released under the Apache 2.0 licence.
- What it gives you: a 0–100 risk score, a label (SAFE, CAUTION or DO NOT INSTALL) and a list of findings.
- Before your first scan: add
--no-llm. Without it, the default setting sends the skill’s file contents to an AI provider. - Where to get it: NVIDIA’s GitHub page only. On 3 Nov 2026 PyPI, the usual Python package store, held no SkillSpector download.
- What it cannot do: it reads files and does not run them, so it misses other languages, text inside images, compiled code and anything that only happens when the skill runs.
Think of the scanner at a building’s door that checks bags for known dangerous shapes. It is quick and catches a lot, but it cannot tell what the person carrying the bag intends, and it does not see what is hidden in a place it cannot scan. A clear result lowers the worry. It does not end it.
Nothing on this page was installed or run by this site. Every behaviour below is described from NVIDIA’s own README and documentation, read on 3 Nov 2026. If that changes, this page will say so with a date.
What is SkillSpector, and who makes it?
The plain point: it is a security checker for AI skills, the folders of written instructions (and sometimes code) that you add to an AI assistant. NVIDIA, the chip company, publishes it.
The README opens with one line: “Security scanner for AI agent skills.”1 The repository sits under NVIDIA’s own organisation account on GitHub, and GitHub lists its licence as Apache 2.0, which allows free use, including at work.2
It was created on 21 March 2026 and had between 23,000 and 25,000 GitHub stars on 3 Nov 2026. Star counts change daily and can be inflated, so they show attention, not quality.3 GitHub shows 21 releases; the newest, v2.13.3, was published on 2 Nov 2026.4 The project file still tags it “Development Status :: 3 - Alpha”, which marks it as early-stage software.5 Expect it to change quickly.
NVIDIA says it uses SkillSpector itself: the README says it “is part of the NVIDIA Verified Skills pipeline, which scans, evaluates, and signs agent skills before publication.”1
SkillSpector is a maintained, free tool from a large company, built for exactly the job of checking a skill before you add it. It is young and alpha-labelled, so treat it as a helper, not a judge.
Why scan a skill at all?
The plain point: a skill can carry code or hidden instructions, and your assistant trusts it. Reading every file takes time, and a scanner gives you a fast first pass.
The README puts the problem this way: skills “execute with implicit trust and minimal vetting.” It also reports that, in the 31,132-skill subset of a research dataset it describes, “26.1% of skills contain vulnerabilities” and “5.2% show likely malicious intent”.1 Those figures come from a research paper the README cites; this site has not opened the paper, so treat the numbers as NVIDIA’s report of it. Other audits are covered on skills are a supply chain.
What does SkillSpector check?
The plain point: it looks for patterns that point to trouble: instructions that try to take over the AI, code that sends data out, and requests for more power than the job needs.
The README says it covers “71 vulnerability patterns across 17 categories”; NVIDIA’s guide page says 68, so the two are out of step.6 The README names the categories: prompt injection, data exfiltration, privilege escalation, supply chain, excessive agency, output handling, system prompt leakage, memory poisoning, tool misuse, rogue agent, anti-refusal, trigger abuse, dangerous code, taint tracking, YARA signatures, MCP least privilege and MCP tool poisoning.6 In plain words:
- Prompt injection is text that tries to give the AI new orders; see what is prompt injection. The README lists “Instruction Override” as “Commands to ignore safety constraints”.7
- Data exfiltration means sending your data somewhere it should not go. The README lists “Credentials (env vars, secrets) flow to network output sinks” as a critical pattern.7
It works in two stages. The first is “Fast static analysis”: it reads the files and matches patterns without running anything. The second is optional: an AI model reads the files and judges intent.6 NVIDIA’s guide explains the difference: static checks are “fast and deterministic”, while the AI stage uses an LLM “to compare what a skill claims with what its code appears to do.”8 An LLM is a large language model, the kind of AI behind chatbots.
How do you install it?
The plain point: install it from NVIDIA’s GitHub page using the command in the README. It needs Python 3.12 or newer, or Docker.
The README’s quick install uses uv, a tool that installs Python programs. uv describes itself as “An extremely fast Python package and project manager, written in Rust.”9 The command, exactly as the README gives it, is below. Install uv first from its own documentation page, and read what its installer does before running it.10
There are other routes in the README: cloning the repository and running make install, or building a Docker image from the included Dockerfile so you need no Python.11 The README also warns: “This project will download and install additional third-party open source software projects. Review the license terms of these open source projects before use.”12
Install it only from NVIDIA’s repository. Check that the address begins github.com/NVIDIA/. On 3 Nov 2026, PyPI’s index listed the name “skillspector” with no files and no versions, and the README’s install command points at GitHub, not PyPI.13 In this site’s reading, that means a command such as pip install skillspector is not the route NVIDIA documents, and anything it installs would not come from NVIDIA’s repo unless NVIDIA says so.
The README’s commands are written for macOS and Linux shells. Its project file lists the operating system as “OS Independent”, but the README does not say whether Windows is supported.5
How do you scan a skill?
The plain point: point skillspector scan at the skill’s folder, file, zip or GitHub address, and add --no-llm so nothing is sent to an AI provider.
The README shows four kinds of target: a folder, a single SKILL.md file, a Git repository address and a zip file.14 The --no-llm flag is listed in its help text as “Skip LLM analysis (static only)”.15
Replace ./my-skill/ with the folder you downloaded. The default report prints in the terminal; add --format markdown --output report.md to save a readable copy.16 Scanning is a read-only look; it is not the same as installing the skill. Download the skill into a folder you do not use for anything else, scan it there, and install only after you have read the report.
Your first scan is skillspector scan ./folder/ --no-llm. It keeps the skill’s contents on your computer, apart from one lookup described below.
How do you read the risk score?
The plain point: the score adds up points for each problem found. Under 21 is labelled SAFE, 21 to 50 is CAUTION, and over 50 is DO NOT INSTALL.
The README gives the arithmetic: each critical finding adds 50 points, high adds 25, medium adds 10 and low adds 5, and executable scripts apply a 1.3 multiplier.17 The scores map to labels like this:
| Score | Severity | Recommendation |
|---|---|---|
| 0–20 | Severity: LOW | Recommendation: SAFE |
| 21–50 | Severity: MEDIUM | Recommendation: CAUTION |
| 51–80 | Severity: HIGH | Recommendation: DO NOT INSTALL |
| 81–100 | Severity: CRITICAL | Recommendation: DO NOT INSTALL |
The program also sets its exit code from the score. That is the number a script or automated check reads, and the README says “the exit code collapses SAFE and CAUTION into 0”, so a scripted check treats CAUTION as a pass unless you ask it to be stricter.18 When you read a report yourself, read the findings, not just the label. A CAUTION report may hold one finding that matters a lot.
The score is a count of flagged patterns, not a measure of how dangerous a skill is. Read the findings behind it.
What does SkillSpector send, and where?
The plain point: by default it sends the skill’s file contents to an AI provider. --no-llm stops that, but a separate lookup still sends the names of any code libraries the skill lists.
The README’s “Trust model and data egress” section is the clearest statement. It gives three facts:
- The AI pass is on by default and sends file contents. “When LLM analysis is enabled (the default), file contents are sent to the active
SKILLSPECTOR_PROVIDERendpoint.” It adds: “Use--no-llmto keep contents local (static analysis only).”19 - OSV.dev gets library names, even with
--no-llm. The supply-chain check “queries OSV.dev with the package names and versions the skill declares”. It “runs even with--no-llm” and sends “dependency coordinates (not file contents)”, with no API key.19 An API key is a password that lets a program use an online service; see what is an API. - It does not run the skill. “It never executes the scanned skill.” The README adds that analysis is “static (regex, Python AST, YARA) plus optional LLM evaluation of file contents”.19
Which AI provider receives the files is your choice. The setting is called SKILLSPECTOR_PROVIDER, and the README says it “Defaults to nv_build”, NVIDIA’s hosted service at build.nvidia.com, which needs an NVIDIA key.20 Other listed providers include OpenAI, Anthropic, Gemini and Ollama.20
Two options need no API key. Ollama is a way to run an AI model on your own computer (see local AI); the README’s example is “Local Ollama — no API key”, with a default address on your own machine (localhost).21 The CLI providers (claude_cli, gemini_cli, opencode_cli) hand the files to a command-line assistant you are already signed in to.1 The README does not say where those assistants send the files, so do not treat them as staying on your machine; only --no-llm is described that way.
The README does not say what happens if you run a plain scan with no key set. Use --no-llm for a first scan and you avoid the question.
Plain skillspector scan sends file contents out. --no-llm keeps them local, but library names still go to OSV.dev. Both facts are NVIDIA’s own.
What can SkillSpector not do?
The plain point: it reads text and code files without running them, so a skill that behaves badly only when it runs, or hides its trick somewhere it cannot read, can pass.
The README’s own limits list: it “May miss patterns in other languages”, “Cannot analyze text in images”, “Cannot analyze compiled or encrypted content”, and is “Static analysis only, no dynamic execution”.22 It also says what it is not: “SkillSpector is defense-in-depth, not a sandbox,” and it “does not contain or isolate a skill you choose to install anyway.”19 A sandbox is a sealed space that stops a program touching the rest of your computer.
Its checks also make mistakes in both directions. The README describes its static stage as “High recall (catches most issues)” but “Moderate precision (some false positives)”, and says the AI stage does not “guarantee a particular precision rate”.23 In plain words, it can flag harmless things, and the README does not claim to catch everything. A false alarm costs you a few minutes. A missed problem is the risk to plan for.
The README makes the same point in its MCP section. An MCP server is a plug-in connection that lets an AI assistant use outside tools; see MCP. That mode reports whether the AI pass ran, “so a low score from a static-only scan is never mistaken for a clean full scan.”24 The same applies to a plain scan with --no-llm: you checked patterns, not intent.
A low score means no flagged pattern was found in the files it could read. It is not proof the skill is safe.
Where does it fit in your checks?
The plain point: run it as a fast first pass, then read the skill yourself with the printable list.
In this site’s reading, a scanner does the boring first sweep, and your own reading catches what a pattern checker cannot: whether the skill asks for more than its job needs, whether the author is who they say, and whether an update changed it. Run the scan again after any update to the skill, because a scan describes the files it read on that day. The steps for who made it, which files to read and the red flags are on check a skill before installing. For where to put a skill once it passes, see how to install AI skills.
Questions people ask
Is SkillSpector free?
The software is free under Apache 2.0.2 The AI pass needs an account or key with whichever provider you choose; check that provider’s own page for terms. The --no-llm scan needs no key (README, above).
Does SkillSpector run the skill it scans?
No. The README says “It never executes the scanned skill” (see the data section above). If you point it at a GitHub address or zip file, it first downloads that content, and the README limits how much it will take in: 100 MiB per download.25
Does a score of 0 mean a skill is safe?
No. A score of 0 to 20 is labelled SAFE, but it only means few or no flagged patterns in the files SkillSpector could read. The limits above still apply, and the README’s labels come from its own scoring rules (see the score section).
Can I install SkillSpector with pip install skillspector?
Not as NVIDIA documents it. The README’s install command uses a GitHub address, and on 3 Nov 2026 PyPI’s index listed no SkillSpector files or versions (sources in the install section). Use NVIDIA’s repository.
Where to go next
- “What is a skill?” See AI skills explained.
- “How do I check one by hand?” Use the printable check a skill before installing list.
- “Why does this matter?” Read skills are a supply chain.
- “How do I add one?” See how to install AI skills.
- “Also adding a connector?” Use before you connect an AI agent or plugin.
- “What do these words mean?” See prompt injection, plugin and repos.
- “Back to the section.” All pages on AI skills.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What does SkillSpector do with a skill it scans?
- It runs the skill in a sealed space and watches what happens
- It reads the skill’s files without running them and gives a risk score from 0 to 100
- It rewrites the skill so any risky lines are removed
Show the answer
Answer: It reads the skill’s files without running them and gives a risk score from 0 to 100. The page says SkillSpector reads the files and never runs the skill. It gives a score from 0 to 100 and a label, and it is not a sandbox.
2. Why does the page suggest adding --no-llm to your first scan?
- It makes the scan find more problems
- It stops the skill’s file contents being sent to an AI provider, which the default setting does
- It lets you scan skills that are not on GitHub
Show the answer
Answer: It stops the skill’s file contents being sent to an AI provider, which the default setting does. By default the AI pass sends file contents to the chosen provider. With --no-llm the contents stay on your computer, although library names still go to OSV.dev.
3. What does a low risk score mean?
- No flagged pattern was found in the files it could read, which is not proof the skill is safe
- The skill has been checked by NVIDIA and approved
- The skill cannot do anything harmful when it runs
Show the answer
Answer: No flagged pattern was found in the files it could read, which is not proof the skill is safe. SkillSpector reads files only. It can miss other languages, text in images, compiled code and behaviour that appears only at run time.
- 1NVIDIA, NVIDIA/SkillSpector README, read at source 3 Nov 2026“Security scanner for AI agent skills.” and “Detect vulnerabilities, malicious patterns, and security risks before installing agent skills.” ↑a“SkillSpector is part of the NVIDIA Verified Skills pipeline, which scans, evaluates, and signs agent skills before publication.” ↑b“AI agent skills (used by Claude Code, Codex CLI, Gemini CLI, etc.) execute with implicit trust and minimal vetting. In the 31,132-skill analyzed subset of the research dataset, 26.1% of skills contain vulnerabilities and 5.2% show likely malicious intent.” ↑c“No API key is needed. Authentication is managed entirely by the agent CLI’s own login session.” ↑d
- 2GitHub API, NVIDIA/SkillSpector repository data, read at source 3 Nov 2026owner “NVIDIA”, type “Organization”; “created_at”: “2026-03-21T00:28:43Z”; license “Apache License 2.0”; “stargazers_count”: 24028; “pushed_at”: “2026-11-02T22:33:56Z” ↑alicense “Apache License 2.0” ↑b
- 3arXiv, Six Million (Suspected) Fake Stars in GitHub: A Growing Spiral of Popularity Contests, Spams, and Malware, v2 6 Sep 2025, read at source 13 Oct 2026 (on /skills/hyperframes/)“the star count is the most widely used popularity signal, but it is also at risk of being artificially inflated (i.e., faked)” ↑
- 4GitHub API, NVIDIA/SkillSpector releases, read at source 3 Nov 202621 releases listed; latest “v2.13.3”, “published_at”: “2026-11-02T06:44:11Z”, “prerelease”: false ↑
- 5NVIDIA, SkillSpector pyproject.toml, read at source 3 Nov 2026classifiers include “Development Status :: 3 - Alpha”; “requires-python”: “>=3.12,<3.16” ↑a“Operating System :: OS Independent” ↑b
- 6NVIDIA, NVIDIA/SkillSpector README, “Features”, read at source 3 Nov 2026“71 vulnerability patterns across 17 categories”; NVIDIA, Scan Agent Skills Before Installation, read at source 3 Nov 2026: “SkillSpector covers 68 vulnerability patterns across 17 categories” ↑a“prompt injection, data exfiltration, privilege escalation, supply chain, excessive agency, output handling, system prompt leakage, memory poisoning, tool misuse, rogue agent, anti-refusal, trigger abuse, dangerous code (AST), taint tracking, YARA signatures, MCP least privilege, and MCP tool poisoning” ↑b“Two-stage analysis: Fast static analysis + optional LLM semantic evaluation” ↑c
- 7NVIDIA, NVIDIA/SkillSpector README, pattern table, read at source 3 Nov 2026“P1 | Instruction Override | HIGH | Commands to ignore safety constraints” ↑a“TT3 | Credential Exfiltration Chain | CRITICAL | Credentials (env vars, secrets) flow to network output sinks” ↑b
- 8NVIDIA, Scan Agent Skills Before Installation, read at source 3 Nov 2026“Static analysis is fast and deterministic.” and “Semantic analysis uses an LLM to compare what a skill claims with what its code appears to do.” ↑
- 9Astral, uv documentation, read at source 3 Nov 2026“An extremely fast Python package and project manager, written in Rust.” ↑
- 10NVIDIA, NVIDIA/SkillSpector README, “Quick install with uv (CLI-only)”, read at source 3 Nov 2026
uv tool install git+https://github.com/NVIDIA/skillspector.git↑ - 11NVIDIA, NVIDIA/SkillSpector README, “From source” and “Docker (no Python required)”, read at source 3 Nov 2026“Run SkillSpector without installing Python by building it locally from the included Dockerfile.” ↑
- 12NVIDIA, NVIDIA/SkillSpector README, “Installation”, read at source 3 Nov 2026“Open-source software notice: This project will download and install additional third-party open source software projects. Review the license terms of these open source projects before use.” ↑
- 13Python Package Index, Simple index for “skillspector”, read at source 3 Nov 2026the listing returned “files”: [] and “versions”: [], with project status “active”. The project’s web page at pypi.org/project/skillspector/ showed a “Client Challenge” page and was not opened past it. ↑
- 14NVIDIA, NVIDIA/SkillSpector README, “Basic Usage”, read at source 3 Nov 2026
skillspector scan ./my-skill/,skillspector scan ./SKILL.md,skillspector scan https://github.com/user/my-skill,skillspector scan ./my-skill.zip↑ - 15NVIDIA, NVIDIA/SkillSpector README, “CLI Options”, read at source 3 Nov 2026
--no-llm Skip LLM analysis (static only); NVIDIA, Scan Agent Skills Before Installation, read at source 3 Nov 2026: under “For static-only review:”,skillspector scan ./my-skill/ --no-llm↑ - 16NVIDIA, NVIDIA/SkillSpector README, “Output Formats”, read at source 3 Nov 2026“Terminal output (default)”;
skillspector scan ./my-skill/ --format markdown --output report.md↑ - 17NVIDIA, NVIDIA/SkillSpector README, “Risk Scoring”, read at source 3 Nov 2026“CRITICAL issues: +50 points”, “HIGH issues: +25 points”, “MEDIUM issues: +10 points”, “LOW issues: +5 points”, “Executable scripts: 1.3x multiplier” ↑
- 18NVIDIA, NVIDIA/SkillSpector README, “Exit codes”, read at source 3 Nov 2026“By default, the exit code collapses
SAFEandCAUTIONinto0. Use--fail-on-findingsto gate on any active finding” ↑ - 19NVIDIA, NVIDIA/SkillSpector README, “Trust model and data egress”, read at source 3 Nov 2026“LLM analysis sends analyzer-eligible file contents to the configured provider. When LLM analysis is enabled (the default), file contents are sent to the active SKILLSPECTOR_PROVIDER endpoint. ... Use --no-llm to keep contents local (static analysis only).” ↑a“SC4 sends dependency names to OSV.dev. The supply-chain check queries OSV.dev with the package names and versions the skill declares, to look up known CVEs. This is fundamental to the check and runs even with --no-llm. It sends dependency coordinates (not file contents), requires no API key, and falls back to a bundled list when OSV.dev is unreachable.” ↑b“It never executes the scanned skill. All analysis is static (regex, Python AST, YARA) plus optional LLM evaluation of file contents — the skill’s code is never run.” ↑c“SkillSpector is defense-in-depth, not a sandbox.” and “It does not sandbox the host. SkillSpector flags risky patterns before you install a skill; it does not contain or isolate a skill you choose to install anyway.” ↑d
- 20NVIDIA, NVIDIA/SkillSpector README, environment-variable table, read at source 3 Nov 2026
SKILLSPECTOR_PROVIDER“Defaults tonv_build.”;NVIDIA_INFERENCE_KEY“Credential for thenv_buildprovider (build.nvidia.com).” “Required for LLM analysis whenSKILLSPECTOR_PROVIDER=nv_build” ↑a“Active LLM provider:openai,anthropic,anthropic_proxy,bedrock,nv_build,gemini,ollama,azure_openai,openai_compatible,claude_cli,gemini_cli, oropencode_cli.” ↑b - 21NVIDIA, NVIDIA/SkillSpector README, “LLM Analysis” examples and provider table, read at source 3 Nov 2026“Local Ollama — no API key”;
ollamaendpointOLLAMA_BASE_URL“(defaulthttp://localhost:11434/v1)” ↑ - 22NVIDIA, NVIDIA/SkillSpector README, “Limitations”, read at source 3 Nov 2026“Non-English content: May miss patterns in other languages”; “Image-based attacks: Cannot analyze text in images”; “Encrypted/binary code: Cannot analyze compiled or encrypted content”; “Runtime behavior: Static analysis only, no dynamic execution” ↑
- 23NVIDIA, NVIDIA/SkillSpector README, “How It Works”, read at source 3 Nov 2026“High recall (catches most issues)”; “Moderate precision (some false positives)”; “Stage 2 does not replace deterministic findings or guarantee a particular precision rate” ↑
- 24NVIDIA, NVIDIA/SkillSpector README, MCP server section, read at source 3 Nov 2026the
scan_skilltool “also reportsllm_used/scan_modeso a low score from a static-only scan is never mistaken for a clean full scan.” ↑ - 25NVIDIA, NVIDIA/SkillSpector README, “Size limits”, read at source 3 Nov 2026“Per-ingest cap:
INGEST_MAX_BYTES(100 MiB) — applied to streamed URL downloads, total uncompressed size of zip archives, and post-clone disk usage of Git repos.” ↑
Read at source 3 Nov 2026: NVIDIA, NVIDIA/SkillSpector (README and pyproject.toml); NVIDIA, Scan Agent Skills Before Installation; GitHub API, repository data and releases; Python Package Index, Simple index for skillspector; Astral, uv documentation. Read at source 13 Oct 2026: arXiv, Six Million (Suspected) Fake Stars in GitHub.
Nothing on this page was installed or run by this site. Every behaviour is described from NVIDIA’s own files, so this page holds no first-hand test results. The door-scanner comparison and the advice marked “this site’s reading” are this site’s own, not claims from these sources. The star count is a snapshot; re-query GitHub rather than trusting it.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK