The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Bash via scripts/gather-evidence.sh and scripts/beautify.sh to fetch and process external plugin code. These scripts are implemented with defensive measures, such as disabling symlinks during git clones (core.symlinks=false) to prevent directory traversal or file read redirections during the audit process.
[REMOTE_CODE_EXECUTION]: The beautify.sh script performs an npm install for the js-beautify package from the official registry. This is a scoped installation within the skill's own directory to provide a pinned version of the tool for the audit task. The gather-evidence.sh script also uses npm view and npm pack to fetch package metadata and tarballs from the official registry for analysis purposes.
[EXTERNAL_DOWNLOADS]: The skill fetches data from registry.npmjs.org (a well-known service) and clones repositories from user-provided URLs. The gather-evidence.sh script restricts URL schemes to https:// or git@ and implements strict handling of the downloaded artifacts, treating them as adversarial data rather than instructions.
[SAFE]: The skill includes explicit instructions for the agent to treat all gathered evidence as data to analyze and never as instructions to follow. It specifically warns against prompt injection attempts within the audited plugins and provides clear guidelines for handling discovered credentials without disclosing their values.

auditing-external-claude-plugins