The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches repository metadata, contributor lists, commit history, and issue data from GitHub's official API to perform the audit layers described in the methodology.
[EXTERNAL_DOWNLOADS]: Downloads model weight files from Hugging Face's official domain to verify parameter counts and perform weight distribution sanity checks. As Hugging Face is a well-known service, these downloads are consistent with the skill's stated purpose.
[REMOTE_CODE_EXECUTION]: Downloads .safetensors files from Hugging Face and processes them using a local inspection script (references/safetensors-inspection.md). While the automated scan flagged this as remote code execution, the external source is a well-known service and the processing script is a static, local component of the skill intended for data validation.
[COMMAND_EXECUTION]: Utilizes shell utilities such as curl, wc, and python3 to retrieve and analyze external model files. These tools are used within the scope of validating the integrity of remote assets.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and analyze untrusted data from third-party GitHub repositories, creating a surface for indirect prompt injection.
Ingestion points: Reads README.md, LICENSE, SECURITY.md, and source code files from external repositories via web_extract and GitHub API calls (SKILL.md).
Boundary markers: The instructions do not define explicit boundary markers or "ignore" directives for the agent when processing the content of audited files.
Capability inventory: The skill has access to execute_code, network operations (via web_extract), and shell command execution.
Sanitization: There is no mention of sanitizing or filtering instructions that might be embedded within the audited repository's documentation or code.

github-project-audit