ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›THE CRAFT›SKILLS ARE A SUPPLY CHAIN
THE CRAFT · METHODADVANCED

Skills Are a Supply Chain

Installing a skill loads someone else’s instructions into your agent. An audit of 3,984 published skills found 13.4% had a critical flaw and 76 were malicious.

READ8 min
WORDS1,539
SECTIONS6
SOURCES1
TYPEGUIDE
CHECKED25 AUG 26
TL;DR — THE SHORT VERSION

Installing a skill is installing someone else's instructions into your agent's context. It is a dependency with none of the tooling that package managers took years to build.

  • A February 2026 audit of 3,984 agent skills found 13.4% contained at least one critical-severity issue.1
  • 76 confirmed malicious payloads built for credential theft, backdoor installation and data exfiltration.
  • Over a third had at least one security flaw of any severity — including prompt injection, text that redirects the agent reading it.
  • There is no lockfile, no signature, no advisory database. You are the review process.
◈ IN PLAIN TERMS

A "skill" is a text file you add to an AI assistant to teach it a task. The assistant reads it and treats it as an instruction.

So installing one from a stranger is like handing someone a note and asking them to do whatever it says. If the note contains something you did not expect, the assistant has no way to know it should not obey.

What you are actually installing

A skill is text that gets loaded into an agent's context and treated as instruction. That is the whole mechanism, and it is why the risk is different in kind from a library.

A malicious library has to execute. A malicious skill only has to be read. Once its text is in context, the agent cannot reliably distinguish it from something you wrote — which is the same structural problem described on guardrails.

The measured position, from a scan of nearly four thousand published skills:2

13.4%
contained at least one critical-severity issue
36.82%
had at least one security flaw of any severity. Snyk’s headline says “Prompt Injection in 36%”; its own table counts this 36.82% as “Skills with any security issue”
76
confirmed malicious payloads, built for credential theft and data exfiltration
ONE SCAN · SKILLS, NOT PERCENTAGES
Snyk’s counts from its February 2026 scan, drawn to the same scale, in the scan’s own wording.
Skills scanned, from ClawHub and skills.sh3,984
At least one security flaw1,467
At least one critical-level issue534
Malicious payloads, confirmed by human review76
Snyk, “ToxicSkills”, 5 Feb 2026, read at source 11 Sep 2026 and quotes re-read 17 Sep 2026: “13.4% of all skills, or 534 in total, all contain at least one critical-level security issue”; “over a third of the ecosystem is affected: 36.82% (1,467 skills) have at least one security flaw”; and “our human-in-the-loop process confirming that 76 of Agent Skills contained malicious payloads”.

What is actually out there

Abstract risk is easy to ignore, so here are five widely-installed skills and what each one is asking for.

And one figure on this list was not a skill at all. "Playwright, 88k" almost certainly refers to Microsoft's Playwright testing framework, not to any Claude skill wrapping it. The skills that do wrap it are small, and there are several. A framework's popularity had been silently transferred to a skill that borrows its name — which is a more useful warning than any star count: check what you are actually installing, not what it is named after.

UI/UX Pro Max — searchable database of UI styles, palettes, font pairings, chart types30k–118k
Taste — adjustable premium defaults, anti-slop, motion-first48k–59k
Impeccable — shared design vocabulary, 23 commands, by Paul Bakaus (creator of jQuery UI)40k–44k
Playwright — opens the app in a browser, screenshots it, flags issuessee below
Huashu-Design — twenty design philosophies, plain MIT since May 2026no dated count

Those are ranges, not numbers, and the ranges are the finding. Verified 22 Aug 2026: UI/UX Pro Max is reported at 29,636 (Snyk, Mar), 94k (Superdesign, Jun) and 118,288 (Skillselion, 20 Aug). Taste at 59.4k in March and 48k in June. Impeccable at 44k in March and 40k in June. Published figures for the same repository differ by two to four times depending on who counted and when — and some counts fall over time, which can mean people removed stars or that the sources counted differently. Recompute against api.github.com before relying on any of them. The Snyk, Superdesign and Skillselion figures were read from those publications; no links were kept. On 23 Sep 2026 GitHub’s API for nextlevelbuilder/ui-ux-pro-max-skill returned “stargazers_count: 129888”.

Read that list again with the permission question in mind. Playwright opens a browser and drives your application. That is a fundamentally different ask from a skill that only shapes text output — and the star count tells you nothing about whether it should have it.

Popularity is a measure of how many people installed it, not of how many read it first.

A correction, made while fact-checking this page. An earlier version of this section said Huashu-Design carried "MIT with a commercial-use condition" and drew a conclusion from it. That was wrong, and it was wrong in an instructive way.

The screenshot it came from showed MIT* with an asterisk and a commercial-licence note. That was the licence until 14 May 2026, when the project moved to plain MIT — commercial use free, no prior authorisation, no fee.3

So the caution stands, but the example inverts: the licence this page was warning about had already been fixed three months earlier. A licence read from a screenshot is a licence at whatever date that screenshot was taken — and licences change in both directions.

What to check before installing one

  • Read it. All of it. A skill is text — usually short. If it is too long to read, that is itself the finding.
  • Does it reference anything outside itself? A URL it fetches, a file it reads, an endpoint it posts to. Each is a channel out.
  • Does it ask for permissions the task does not need? A formatting skill needing network access is the clearest signal there is.
  • Does it contain instructions about instructions? "Ignore previous", "always", "regardless of" — language aimed at the agent's obedience rather than the task.
  • Who publishes it, and does the account have history? A skill published last week by an account with nothing else is not a package with a maintainer.

The printable, file-by-file version of this list, with the red flags from the incidents below: check a skill before installing.

A malicious library has to execute. A malicious skill only has to be read.

What the ecosystem does not have

Worth naming plainly, because the absence is easy to miss when installation is one click:

  • No lockfile. A skill can change under you between runs.
  • No signature. Nothing proves the author is who the listing says.
  • No advisory database. When a skill is found malicious, there is no channel that tells you.
  • No dependency graph. A skill invoking another skill is invisible until it happens.

npm took years to build those. Skills are at year one and moving faster.

What has happened since: malware, a ransomware test and scanning

Three things since this page was first written show the risk is not theoretical, and what the first official fix looks like.

Hundreds of malicious skills on one marketplace. In February 2026, Koi Security audited ClawHub, the skill marketplace for the OpenClaw agent, and reported “341 malicious skills” out of 2,857. Most of them, 335, told users to install fake prerequisites that were really a data-stealing program for Macs. The only bar to publishing was a GitHub account at least a week old.4 ClawHub serves OpenClaw, not Claude, but the format is the same open skill format, and so is the trick: the skill text looks harmless and the install step does the damage.

A ransomware proof of concept with Anthropic’s own skill. In December 2025, a researcher at Cato Networks took Anthropic’s open-source GIF Creator skill and added “a seemingly benign function that downloads and executes outside code”, and showed the skill could be made to run ransomware. Claude checked the code it could see, but “had no visibility into the outside code the tool downloaded later”. Anthropic’s reply was that “it is the user’s responsibility to only use and execute trusted Skills.”5 This was a controlled test, not an attack seen in the wild. Its lesson is the one above: reading the skill is not enough if the skill fetches more code later.

The first official scanning. Anthropic now scans third-party skills and plugins for malicious content on its Enterprise plans when someone uploads or edits them. The help centre says it is “off by default until October 2, 2026, when it turns on for Enterprise organizations that haven’t set it.”6 It does not cover everything: skills uploaded through the API are not scanned.7

The working rule

Treat a skill as you would a script from a stranger, because that is exactly what it is. Read it, run it at tier one first, and check what it did against what it claimed.

And pin the ones you rely on. A skill you have read is a skill you have read at one version — nothing guarantees the next one is the same.

Related on this site: ComfyUI Explained · Before You Connect an AI Agent or Plugin · How to Install AI Skills in Claude and Other AI Apps · AI Skills: Install, Compare and Check Them · We Checked the Five ‘Claude Video Skills’ From a Viral Reel · What Is SkillSpector?

SOURCES
  1. 1Snyk, “ToxicSkills”, 5 Feb 2026, quote re-read 22 Sep 2026“13.4% of all skills, or 534 in total, all contain at least one critical-level security issue”. ↑
  2. 2Snyk, "ToxicSkills", 5 Feb 2026, read at source 11 Sep 2026 and quotes re-read 17 Sep 2026“13.4% of all skills, or 534 in total, all contain at least one critical-level security issue” across 3,984 skills scanned from ClawHub and skills.sh; “over a third of the ecosystem is affected: 36.82% (1,467 skills) have at least one security flaw”, and “our human-in-the-loop process confirming that 76 of Agent Skills contained malicious payloads”. — 3,984 skills scanned ↑
  3. 3alchaincyf/huashu-design README and LICENSE, checked 22 Aug 2026↑
  4. 4The Hacker News, Researchers Find 341 Malicious ClawHub Skills Stealing Data from OpenClaw Users, 2 Feb 2026, read at source 13 Oct 2026“A security audit of 2,857 skills on ClawHub has found 341 malicious skills across multiple campaigns, according to new findings from Koi Security”; “335 skills use fake pre-requisites to install an Apple macOS stealer named Atomic Stealer (AMOS).”; “The only restriction at this stage is that a publisher must have a GitHub account that’s at least one week old.” Koi Security’s own post was not read. ↑
  5. 5Axios, report on Cato Networks’ research, 2 Dec 2025, read at source 13 Oct 2026 (in a browser; plain fetches returned 403): “An AI tool that Claude uses to automate tasks can be easily weaponized to execute ransomware”; “Cherny modified the widely shared, open-source "GIF Creator" skill built by Anthropic by inserting a seemingly benign function that downloads and executes outside code.”; “Claude reviewed the visible code inside the Skill before running it, but it had no visibility into the outside code the tool downloaded later”; Anthropic: “it is the user’s responsibility to only use and execute trusted Skills.” Cato’s own post sits behind a bot check and was not read.↑
  6. 6Claude Help Center, Get started with skill and plugin scanning, read at source 13 Oct 2026“Skill and plugin scanning automatically checks third-party skills and plugins for malicious content when someone uploads or edits them”; it is “available on Enterprise plans”; “Skill and plugin scanning is off by default until October 2, 2026, when it turns on for Enterprise organizations that haven’t set it.” ↑
  7. 7Anthropic, Agent Skills overview, read at source 13 Oct 2026“Scanning doesn’t cover Skills uploaded through the Skills API or the Claude Console.” ↑

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
◈ IF YOU ARE CITING THIS

Cite the original source, not this page. Every figure here names the organisation that issued it and the date it was published — those are the citations worth carrying. This page is a signpost, not a primary source.

If you need to reference the collation itself — the comparison, the framing, or a correction made here — the press page has the details. But if you are quoting a number, go to whoever measured it.

Or check it yourself. How to check the figures here names the feed or document behind each recurring source, and what to expect when your number differs from ours.

aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE