- A skill can run code on your computer. Installing one is installing software, so check it like software.
- Read every file, not just the description. The risky part is often a script or an install step.
- The checking is your job. After a researcher showed a modified skill could deliver ransomware, Anthropic replied that “it is the user’s responsibility to only use and execute trusted Skills.”1
Before you install
- ☐ Who made it? Prefer skills you wrote or got from Anthropic or the tool’s own maker. 2
- ☐ Install from the official repo only, the maker’s own GitHub page, not a copy on a directory site. A marketplace name tells you who runs the catalogue, not what each item does. 3
- ☐ Popular is not checked. Snyk scanned 3,984 published skills and found a critical issue in 13.4% and 76 malicious payloads. 4 Koi Security found 341 malicious skills among 2,857 on ClawHub, the marketplace for the OpenClaw agent. 5
- ☐ Plugin or skill? A plugin can bundle skills, scripts and connectors, and runs with your permissions. 3
- ☐ Read the licence file. No licence file means no stated permission to copy or reuse it.
- ☐ Optional: scan it first. NVIDIA’s free SkillSpector gives a skill a 0–100 risk score. Add
--no-llmso the skill’s contents stay on your computer. A low score is not proof of safety.6
Read these files
- ☐ All of it, not just SKILL.md. Anthropic says to review “SKILL.md, scripts, images, and other resources.” 2
- ☐ The allowed-tools line at the top of SKILL.md: it can switch on tools without asking you. 7
- ☐ Every web address it reads from or sends to. 2
- ☐ The install steps in the README, especially any command that downloads and runs something.
Red flags: stop here
- ☐ “First install this” prerequisites from a link. Koi found 335 skills using fake prerequisites to install malware that steals data from Macs. 5
- ☐ Code that fetches more code later. In Cato Networks’ test, Claude read the visible code but not what it downloaded afterwards. 8
- ☐ Network calls or file reads the job does not need. 2
- ☐ Lines aimed at the AI’s obedience: “ignore previous”, “always”, “regardless of”, or orders to connect to unknown servers. 9
- ☐ Passwords or keys written into the files, or a request for yours. 10
- ☐ Too long to read. That is the finding.
After you install
- ☐ Turn telemetry off (usage data sent to the maker) wherever there is a switch.
- ☐ First run in a test folder with no private files, and approve each command.
- ☐ Re-read after updates; leave auto-update off for skills you rely on. 2
- ☐ Remove skills you no longer use.
- ☐ At work, ask about scanning. Claude Enterprise checks third-party skills on upload, on by default from 2 Oct 2026 unless an owner chose otherwise; it does not cover skills sent through the API. 11
Who made it, what every file does, and what it can reach. If you cannot answer all three, do not install it.
Related on this site: What Is a Plugin in AI?
- 1Axios, report on Cato Networks’ skill ransomware test, 2 Dec 2025, read at source 13 Oct 2026 (in a browser; plain fetches were refused): Anthropic said “it is the user’s responsibility to only use and execute trusted Skills.”↑
- 2Anthropic, Agent Skills overview, read at source 13 Oct 2026“Use Skills only from trusted sources: those you created yourself or obtained from Anthropic.” ↑a“Review all files bundled in the Skill: SKILL.md, scripts, images, and other resources.” “Treat like installing software”. ↑b“Skills that fetch data from external URLs pose particular risk”. ↑clook for “unexpected network calls, file access patterns, or operations that don’t match the Skill’s stated purpose”. ↑d“Even trustworthy Skills can be compromised if their external dependencies change over time”. ↑e
- 3Claude Code Docs, Plugin security, read at source 13 Oct 2026“A marketplace’s name tells you who publishes the catalog, not what each plugin in it does”. ↑a“A Claude Code plugin you install can execute arbitrary code on your machine with your user privileges.” ↑b
- 4Snyk, “ToxicSkills”, 5 Feb 2026, read at source 13 Oct 2026“scanning 3,984 skills from ClawHub and skills.sh as of February 5th, 2026”; “13.4% of all skills, or 534 in total, all contain at least one critical-level security issue”; “76 malicious payloads designed for credential theft, backdoor installation, and data exfiltration.” ↑
- 5The Hacker News, Researchers Find 341 Malicious ClawHub Skills, 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”. ↑a“335 skills use fake pre-requisites to install an Apple macOS stealer named Atomic Stealer (AMOS).” ↑b
- 6NVIDIA, NVIDIA/SkillSpector README, read at source 3 Nov 2026“Risk scoring: 0-100 score with severity labels and clear recommendations”; “Use --no-llm to keep contents local (static analysis only).” ↑
- 7Claude Code Docs, Skills, read at source 13 Oct 2026“A skill can grant itself broad tool access, so review the allowed-tools of skills checked into a repository before you run Claude Code there.” ↑
- 8Axios, report on Cato Networks’ skill ransomware test, 2 Dec 2025, read at source 13 Oct 2026“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”. A researcher’s controlled test, not an attack seen in the wild. ↑
- 9Claude Help Center, Use skills in Claude, read at source 13 Oct 2026“pay attention to instructions or code within the skill that instruct Claude to connect to potentially untrusted external network sources.” ↑
- 10Claude Help Center, How to create custom skills, read at source 13 Oct 2026“Don’t hardcode sensitive information (API keys, passwords).” ↑
- 11Claude 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”; “Skill and plugin scanning is off by default until October 2, 2026, when it turns on for Enterprise organizations that haven’t set it.” Anthropic, Agent Skills overview, read at source 13 Oct 2026: “Scanning doesn’t cover Skills uploaded through the Skills API or the Claude Console.” ↑
Anthropic, Agent Skills overview; Claude Code Docs, Skills and Plugin security; Claude Help Center, Use skills in Claude, How to create custom skills and Get started with skill and plugin scanning; all read at source 13 Oct 2026.
Snyk, “ToxicSkills”, 5 Feb 2026; The Hacker News, Researchers Find 341 Malicious ClawHub Skills, 2 Feb 2026 (Koi Security’s findings, as reported; Koi’s own post was not read); Axios, Cato Networks skill ransomware report, 2 Dec 2025 (Cato’s own post was not read: it sits behind a bot check). All read at source 13 Oct 2026.
The grouping into four steps, the licence and telemetry ticks, and the three-question takeaway are this site’s own method, from the pages linked above; they are not claims from these sources.
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