skill-installer
Warn
Audited by Gen Agent Trust Hub on Sep 18, 2026
Risk Level: MEDIUMCREDENTIALS_UNSAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill downloads content from external GitHub repositories. While it defaults to curated skills from a trusted organization, it explicitly allows the agent to install skills from any user-provided GitHub repository owner and path.
- [REMOTE_CODE_EXECUTION]: The skill's primary function is to download and install executable logic in the form of agent skills. While the installation script itself does not execute the newly downloaded code, the purpose of the installer is to facilitate the addition of new executable components to the agent platform.
- [COMMAND_EXECUTION]: The installer script uses subprocess.run to execute git commands for repository cloning and sparse checkouts. The arguments for these commands are constructed from repository information provided to the skill at runtime.
- [CREDENTIALS_UNSAFE]: The scripts check for and utilize GITHUB_TOKEN or GH_TOKEN environment variables to authenticate with GitHub. While these tokens are sent to official GitHub domains (api.github.com and codeload.github.com), they provide access to private repository content and represent sensitive environment data.
- [INDIRECT_PROMPT_INJECTION]: The skill serves as a vector for indirect prompt injection by ingesting instructions from external repositories. A malicious repository could contain instructions in its SKILL.md files designed to subvert agent behavior.
- Ingestion points: Files downloaded from GitHub (specifically SKILL.md and scripts) in scripts/install-skill-from-github.py.
- Boundary markers: None identified to isolate or warn the agent about the untrusted nature of the downloaded instructions during installation.
- Capability inventory: Subprocess execution for git commands, file system writes via shutil.copytree, directory creation via os.makedirs, and network access via urllib.
- Sanitization: While path traversal checks (os.path.realpath and relative path validation) are performed on filenames, the textual content of the installed skills is not inspected or sanitized.
Audit Metadata