Should we ban AI skills from the internet?

No. Many are useful. Check each one for risky behaviour before it is used, have an accountable person approve it, and keep the approved version in one place so nobody needs to copy it from elsewhere.

How do we find out which AI skills people already use?

Ask teams to bring the skills they rely on into one shared library, and make that library the easiest place to get approved skills. People share what they use when sharing is easy and useful to them.

Who should approve an AI skill?

It depends on the risk. A low-risk skill can be approved by its owner. A skill that handles client data or makes decisions with financial impact should also be approved by an admin, the security team, or two people.

Is testing every skill a lot of work?

Not compared with the cost of a skill that is wrong for everyone who uses it. A handful of real cases with known results, a few colleagues trying it, and a run in each AI tool is a small effort next to fixing the work of everyone who used a faulty skill.

ResourcesOpinion

Your company is sharing AI skills nobody has tested

AI skills spread through companies the way good ideas do: person to person, tool to tool. What usually does not travel with them is any testing, approval or record of where they came from.

The Fygment team

25 October 2026, 5 min read

Here is how most AI skills reach a team. Someone gets good results from a set of instructions, saves them as a skill, and posts it in a chat channel. A colleague copies it into Claude, another into ChatGPT, a third changes a line to suit their own work. A month later there are four versions in use, nobody knows which one is right, and the only testing any of them had was the author trying it twice.

That is how good ideas spread, and it is a sign that people are finding real value in AI. The problem is what does not spread with the skill: any testing, any approval, and any record of where it came from.

Skills move faster than anyone can see

AI use at work is already running ahead of oversight. In a 2025 survey reported by Infosecurity Magazine, 69% of cybersecurity leaders told Gartner they had evidence of, or suspected, employees using public AI tools. In 1Password's 2025 report, as covered by Expert Insights, a quarter of employees said they use AI apps at work without IT oversight.

Skills follow the same path, with one difference: a skill does not just answer a question, it shapes how the AI does a piece of work, every time someone uses it. When the skill is wrong, the work is wrong in the same way for everyone who copied it.

Working twice is not the same as tested

Most shared skills have been tested by exactly one person, on one or two tasks, in one AI tool. AI tools answer a little differently every time, so a skill that worked twice can still fail often for the next person. Anthropic's own checklist asks for at least three test cases, tested on real tasks and with feedback from colleagues, before a skill is shared. We explain why in AI skills aren't one-and-done.

A tested skill has been run on a handful of real cases with known correct results, used by a few colleagues on their own work, and tried in each AI tool it will be used in. That is not a heavy process. It takes far less time than finding out later, from colleagues, that the skill was wrong, and it is the difference between a skill people can rely on and one they have to double-check.

“A skill that reaches the whole company should have passed more than its author's two tries.”

Skills from outside need checking first

Some of the skills circulating in a company were never written there. Public skill directories are growing fast, and most of what they hold is shared in good faith. But they are also a new way in for bad actors. When Snyk audited nearly 4,000 public skills, it found 13.4% with at least one critical security issue and 76 with confirmed malicious payloads. A separate audit by Koi Security found 341 malicious skills among 2,857 in one directory, many hidden behind professional-looking instructions. OWASP, the open security community, ranks malicious skills as the number one risk for agent skills.

None of this means outside skills should be banned. It means they need the same treatment as any other software a company brings in: checked before anyone uses them, and approved by someone accountable.

Why banning does not work

The reflex is to forbid skills that were not built by IT. Gartner advises against it: in its 2026 guidance on AI agent sprawl, it warns that when companies block agents, employees tend to go around the controls and turn to riskier shadow AI. The same release reports that only 13% of organisations believe they have the right governance for AI agents in place.

The better answer is to make the safe path the easy path. If the tested, approved version of a skill is the one people find first, in the AI tools they already use, there is little reason to copy an untested one from a chat message.

What a proper path looks like

Four things are enough to start. Every skill that is shared has been tested on real cases. Every skill from outside is checked before it comes in. Every shared skill has an owner and has been approved by the right person. And everyone gets their skills from one company library, in every AI tool they use, so there is one current version instead of four. We set out the full model in How to govern AI skills across a company.

This is what Fygment does. Outside skills are scanned for risky behaviour before they can be imported. Each skill has an owner, test cases a new version must pass, and an approval step before it reaches anyone. The approved version is delivered to Claude, ChatGPT, Copilot and the other AI tools your team uses, and every change is recorded.

If you want to give your team's skills a proper path, Fygment is free to start.

Questions we often get

Should we ban AI skills from the internet?: No. Many are useful. Check each one for risky behaviour before it is used, have an accountable person approve it, and keep the approved version in one place so nobody needs to copy it from elsewhere.
How do we find out which AI skills people already use?: Ask teams to bring the skills they rely on into one shared library, and make that library the easiest place to get approved skills. People share what they use when sharing is easy and useful to them.
Who should approve an AI skill?: It depends on the risk. A low-risk skill can be approved by its owner. A skill that handles client data or makes decisions with financial impact should also be approved by an admin, the security team, or two people.
Is testing every skill a lot of work?: Not compared with the cost of a skill that is wrong for everyone who uses it. A handful of real cases with known results, a few colleagues trying it, and a run in each AI tool is a small effort next to fixing the work of everyone who used a faulty skill.

Keep reading

How to govern AI skills across a company: owners, approvals and review cycles
Governing AI skills does not need a committee. It needs eight things: know what you have, give every skill an owner, set the rules for who may create and share, check outside skills, approve by risk, roll out the approved version to every AI tool, review on a cycle, and see what is actually used.
AI skills aren't one-and-done: the mistakes companies make
Most AI skills are written once, tried twice and shared with the whole team. These are the six mistakes we see most often, why they happen, and what the teams that get skills right do differently.