Last updated: 2026-11-20
The short answer
No scan can tell you that a skill is harmless. A scan can give you a reading list: which parts of the skill do something you should look at before you let your agent use it.
When you install a skill into a local agent such as Claude Code, Codex, or Cursor, its instructions steer your agent, and any scripts it ships run on your machine with your agent's permissions. So the question to ask is not "did it pass?" but "does what it does match what it says it is for?"
This article explains how to read the scan notes on a SourceWeft skill page, what each flag means, a worked example on a real skill, and what you still have to check yourself.
What a skill can do to you
A skill is a folder with a SKILL.md file, and sometimes scripts, reference files, and assets. Two things matter for risk:
- The instructions are executed by a model. Text in
SKILL.mdcan tell your agent to read files, run commands, or send data somewhere. A skill does not need a script to cause harm. - Scripts are code. If a skill ships scripts and you install it for a local agent, those scripts can run on your computer.
If you want the general steps for finding and installing skills, see where to find agent skills and how to install them. This article is about the checking step.
What the SourceWeft scan does
When a skill is submitted to the SourceWeft directory, an automated scan reads the files in the skill bundle and looks for patterns. It is a static pattern sweep: it does not run the skill. For files it cannot read as text, the only question it asks is whether they are executable code.
There are three possible results:
- Refused. If the scan finds a pattern that is the attack itself, the submission is not stored, and the submitter is told why. There are three such patterns: fetching remote code and piping it into a shell (
curl | sh), decoding and running an encoded command (base64 … | sh), and text that tries to override the assistant's instructions. - Published with notes. Flags that describe what a skill does, rather than that it is hostile, are recorded and shown on the skill's page under Automated scan notes. The page says: "Our automated scan flagged the following in this version. Flags are advisory — read the source before installing." A skill that only carries these flags is published and listed without waiting for a person.
- Held for an admin. A skill that ships a compiled binary waits for an admin before it is listed. A flag the scan does not recognize also waits as a draft.
The point of the second result is that a flag is information, not a verdict. Many legitimate skills do things the scan notices. A skill that teaches an API will mention credentials and make network calls.
Because flagged skills can be listed without a person reading them, the flags are for you. They tell you where to look.
The flags and the question each one raises
These are the labels shown on the skill page.
| Flag as shown on the page | What was found | What to check |
|---|---|---|
| Makes outbound network calls | Code that references an HTTP library or a network-request call | Where does it connect, and does SKILL.md say why? The flag can come from an import alone. |
| Posts data to an external host | A post(…) call, a fetch(…) to a URL, or curl -X POST | Is it actually sending data, and what data, to whom? A plain fetch of a URL can trigger this flag too. This is the one to read most closely. |
| Reads environment variables | A script or instruction that reads env vars | Which variables? Its own settings, or things like API keys? |
| Reads credential files | A pattern that reads credential-like files | Which files, and does the skill's job need them? |
| Reads another skill's files | A path into a different skill's folder | Why does this skill depend on another one? |
| Requests code-execution tools | The skill asks for code-execution tools | Which commands, and are they limited? |
| Mentions the system prompt | The text refers to the system prompt | Is it documentation, or an instruction to change behavior? |
| Ships scripts its instructions never mention | Scripts that SKILL.md does not describe | Read every script; the instructions will not tell you what they do. |
| A script reaches files outside the skill | A script uses paths outside the skill folder | Which paths? A cache directory is different from your home directory's contents. |
| Ships an executable binary | A compiled or opaque file | You cannot read it. Install only if you trust the publisher. |
New submissions that match one of the three refused patterns are not stored, so they will not appear as flags on a newly listed skill. Versions that were added before this check existed are not covered by that statement, so still read what you install.
A worked example on a real skill
The skill gke-kueue-troubleshooting is published in the directory and carries three flags. We installed it into an empty folder with the published CLI (version 0.3.0) and read its scripts. This shows what the flags look like in practice. A flag on a skill is not a judgment on its author.
On the skill page, under Automated scan notes, the version we looked at lists:
- Makes outbound network calls
- A script reaches files outside the skill
- Reads environment variables
Running npx @sourceweft/cli skills info @google/gke-kueue-troubleshooting shows the same flags by their identifiers:
Scripts: yes — this skill includes executable files
Scan flags: egress:fetch, script:out-of-bundle-path, secrets:env-access
What we found when we read the scripts at that commit:
- Outbound network calls:
kueue_schema.pydownloads documentation and release files fromkueue.sigs.k8s.ioand from GitHub's releases. The destination is a project documentation site, and the same behavior can be switched off with an environment variable. - Environment variables: The scripts read variables that are the skill's own settings, such as a cache directory and a switch that turns off the remote download. They also read the standard
XDG_CACHE_HOMEto find a cache location. - Files outside the skill: The scripts write a cache to a per-user folder,
~/.cache/gke-kueue-troubleshooting, by design "never inside the skill directory", as a code comment puts it.
Each flag has a concrete explanation, and each can be checked in a few minutes. That is the use of a flag: it tells you which lines to read. For a different skill, the same three flags could have a worse explanation. You would only learn that by reading.
A checklist you can run in five minutes
- Look at the trust label. The CLI prints one of four. "Verified" means a SourceWeft admin reviewed it. "Featured" means a featured publisher, not reviewed by an admin. "Featured, Verified" means a featured publisher whose skill an admin also reviewed. "Community" means not reviewed by SourceWeft. For "Featured" and "Community", no SourceWeft admin has reviewed the skill, so the review is yours.
- Run
skills infobefore you install. It prints the source repository, the exact commit, the license, whether scripts are included, and the scan flags. - Read
SKILL.md. Check that its instructions are ones you would give your agent yourself. - Read the scripts, if there are any. Start with the ones that match a flag. Search for the host names and environment variable names.
- Check the source. Is the repository from the publisher you expect? Is the commit the one you reviewed?
- Prefer a project-scoped install first.
--scope projectinstalls into the current project rather than your home directory, so a skill you are still evaluating does not apply everywhere.
What the scan cannot do
- It does not run the skill. It matches patterns in the text. A script that builds a command at run time can avoid a pattern.
- It matches text, including documentation. A phrase such as
.env,process.env, or "system prompt" inSKILL.mdor a reference file can raise a flag even when nothing is read or run. Some flags are explained by a sentence in the docs; open the file and line before you decide. - It cannot read a binary. It can only say that a file looks like code.
- It does not know your context. A skill that reads environment variables is fine in one setup and a problem in another.
- It does not review the publisher. It reads files, not people.
- A clean result is not a guarantee. It means the scan found nothing in the patterns it checks.
If a skill matters to you, read it. The scan is there to make that faster.
FAQ
Are agent skills safe to install?
Skills are third-party instructions, and some include scripts. Whether one is appropriate for you depends on what it does. Read SKILL.md and any scripts, and use the scan notes to decide where to look first.
What does "Automated scan notes" mean on a skill page? It lists what an automated pattern scan found in that version of the skill, for example that a script makes network calls. The page says the flags are advisory, so you should read the source before installing.
Does a flag mean the skill is malicious? No. The flags describe behavior, and many legitimate skills trigger them. For new submissions, the patterns that indicate an attack itself are refused at submission and not stored.
Does a skill with no flags have no risk? No. A scan checks a fixed set of patterns in text. It cannot tell whether a skill's instructions are appropriate for you.
Who reviews skills in the directory? Only skills labelled "Verified" were reviewed by a SourceWeft admin. Others are scanned automatically and are not reviewed by a person.
Can I see the flags before I install?
Yes. The skill page shows them under "Automated scan notes", and npx @sourceweft/cli skills info <skill> prints them before you install anything.