Hands
NL
Sign in Get started with Hands

An AI skill is not just a harmless prompt

AI skills are not automatically safe. Understand the risks, what the Hands skill scan checks, and why a scan is not a security guarantee.

Explainer · 7 min

Imagine finding a skill that turns customer conversations into perfect meeting notes. A clear description, enthusiastic users, and a useful example. Exactly what you need.

During setup, your AI just needs to download a helper program. Then the conversation files need to go to an external service for analysis. That could be a legitimate workflow. But those are two decisions that go well beyond “Use this format for my notes.”

The question is not only whether a skill works. It is what you are asking your AI to do with it.

Instructions and access

An AI skill is a package of working instructions, often in a file called SKILL.md. It can also include templates, reference material, and executable scripts. An agent loads those resources when a task needs them. 1

A skill does not, by itself, give your AI access to your entire computer, inbox, or cloud storage. Access depends on the tools, connections, and permissions available in the environment. The risk lies in how the skill directs the agent to use those existing capabilities. 2

Summarizing a document is different from sending the entire project folder to an unknown server. When an agent can both read files and send data, that distinction matters. 3

Text is the packaging. Not the security boundary.

Assess a skill as a workflow that can direct real actions, not simply as a piece of text.

Risks beyond the description

A skill can direct an agent to download malicious software, request secret keys, or send data elsewhere. Ordinary language can carry an attack too: a hidden instruction to disregard earlier rules or conceal an action from the user. This is called prompt injection. 4

The problem can also arrive later. A skill that looks understandable at installation can fetch new code or instructions during use. The content behind that external reference can change. 2

That does not make every download or script malicious. Commands such as pip install or npm install can be entirely normal. The useful questions are: where does the package come from, what will it do, and does this task need it? Recognizing a familiar command is not a review of the whole workflow. 3

What the research shows

In Agent Skills in the Wild, published in January 2026, researchers analyzed 31,132 skills. They reported at least one vulnerability in 26.1%. In 5.2%, they identified high-severity patterns that their method considered strongly indicative of malicious intent. That is not the same as individually confirming every one of those skills as malware. 5

Snyk studied a different collection of 3,984 skills in February 2026. It reported a security issue of some severity in 36.82%. Its researchers also confirmed 76 malicious skills through human review of the investigated data. 6

Those percentages belong to particular datasets, dates, and methods. They are not a current probability estimate for every skill you encounter, and they do not describe the proportion of unsafe skills in Hands. The practical lesson is simpler: popularity and a polished description are not evidence of safety.

The skill scan in Hands

On a Hand where scanning is available, Chat → Skills shows a shield with a magnifying glass beside the skill. Its label is “Check skill safety.” The icon starts an action; it is not a badge saying the skill has already been found safe.

Start a check. Do not read it as a guarantee.

This is the same symbol used for the scan action in the console. It deliberately has no checkmark suggesting safety before a scan.

Hands uses SkillSpector, NVIDIA’s open-source scanner. It combines static checks with optional AI-based analysis of instructions and code. The Hands integration uses both. 7

The scan works on a temporary copy of the skill files. It does not run the inspected scripts to see what happens. You receive progress updates, findings, a risk score, and a detailed report. The scan does not modify the original skill. AI review uses the platform; this is not a promise that all analysis happens exclusively locally.

A higher score means a higher assessed level of risk. It is not a percentage predicting how likely an attack is. The findings and their explanations matter more than the number alone.

Current behavior — checked on November 13, 2026

The button starts a manual review of a skill that is already present. It is not yet an automatic security check before every upload. The scan does not automatically remove, repair, or block the skill. Availability depends on your Hand.

This also explains why the icon can appear on built-in skills: the current scan action does not distinguish between built-in and personal skills. The button does not mean suspicious content has been found.

What a scan cannot guarantee

A scanner can miss something. A warning can also concern behavior that is legitimate in the right context. In June 2026, Trail of Bits demonstrated bypasses of several skill scanners. Its warning was to avoid outsourcing trust entirely to a scanner. That research is not an evaluation of this Hands integration. 8

So “no findings” does not mean “proven safe.” An incomplete or failed scan does not provide a safety verdict. The Hands integration makes that distinction explicit.

A review cannot guarantee the safety of files that change later or extra code a skill subsequently fetches. Source selection, limited access, and attention to updates still matter. 2

Four questions before use

Who supplied this workflow? Choose a traceable source and examine its maintenance history. A known supplier is a better starting point than an anonymous download, but it does not replace reviewing the version you actually use. 4

Does its behavior fit the task? Review supporting files, scripts, and external references too. A formatting skill has a different job from an email assistant. Access that does not fit the task needs an explanation. 3

What could the agent reach? Limit files, connections, and privileges to what is needed. Do not try unfamiliar skills with customer data or production keys. Keep secrets out of skill text. An isolated workspace does not automatically protect the services you connect to it. 3

What could change tomorrow? Review updates and additional downloads. An assessment of version A is not approval for version B. The source, installed files, and assessment need to remain tied together. 8

Next: review during upload

The natural place for this review is when you add a skill. That is the direction in which we want to improve the process:

Upload → keep separate → scan → review findings → deliberately add or cancel.

The important difference is that the skill would not become available first and be reviewed afterwards. The files ultimately added must be the same files that were checked. A warning or unfinished scan must not disappear behind a success message.

For built-in skills, that responsibility belongs with Hands. We want to organize review of the supplied version rather than ask every customer to repeat the same check. This is a development direction, not a claim that this upload flow is already available or that all built-in skills are centrally certified.

More capability, less blind trust

Skills make good workflows easier to reuse. We want to preserve that benefit: not by treating every script as suspicious, but by making the source, requested actions, and appropriate access clear.

Use the scan as additional information. Read the findings. And start with one question whenever you consider a new skill:

The question that matters

“If this skill did something other than what it promises, what could it reach with the access my AI has?”

A good skill makes your AI more useful. A good review helps you decide under which conditions to use it.

Sources

The research findings below refer to the stated publication dates and datasets. Product behavior in this article was checked on November 13, 2026.

  1. Agent Skills — the open format and skill contents.
  2. Anthropic — Agent Skills: operation and security considerations.
  3. Anthropic — Skills for enterprise: review and access risks.
  4. ESET Research — observed abuse of AI skills, March 20, 2026.
  5. Liu et al. — Agent Skills in the Wild, January 15, 2026.
  6. Snyk — ToxicSkills research, February 5, 2026.
  7. NVIDIA — SkillSpector: static and semantic analysis.
  8. Trail of Bits — The sorry state of skill distribution, June 3, 2026.

Know your skill. Stay in control.

Learn how to work with skills and choose which workflows your Hand can use.

← Back to all Insights