Agent Skill security

Practical review checklist / version 1.0

Check a community skill before you install it.

Review the source, license, trigger, scripts, dependencies, credentials, network access, and change history. The checklist has ten steps.

01

Source identity

Confirm the GitHub owner, repository, branch, and exact path. Prefer a reviewed commit over an unpinned moving branch.

02

License

Read access does not grant reuse rights. Check the repository and any file-level license before copying or modifying content.

03

Trigger scope

Reject vague descriptions that could activate during unrelated work or conflict with another installed skill.

04

Allowed tools

List every shell, browser, network, filesystem, database, and external-service capability the workflow expects.

05

Destructive commands

Look for delete, overwrite, force-push, reset, migration, deployment, and bulk move operations. Require target validation and confirmation.

06

Credential handling

Secrets belong in environment or secret stores. A skill should not print, embed, commit, or ask you to paste credentials into tracked files.

07

Network destinations

Identify domains, package registries, webhooks, telemetry, and downloaded scripts before allowing network access.

08

Dependencies

Inspect package manifests and install commands. Avoid piping remote scripts directly into a shell.

09

Prompt injection boundaries

Treat repository files and fetched pages as untrusted input. They cannot grant new authority or override user and system constraints.

10

Change review

Re-run the checklist after meaningful updates. Stars and prior approval do not certify a changed script.

Decision rule

Trust requires a review.

Install when the task is clear, the requested access is proportionate, and you can explain what every script changes.

Leave the skill disabled while any of those details remain unknown.