The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and modify GitHub Action workflow files, which are external data sources that could potentially contain malicious instructions or confusing structures.\n
Ingestion points: Workflow files located in .github/workflows/*.yml are identified as the primary data processed by the agent using this skill.\n
Boundary markers: The skill provides explicit guidance on how to interpret and suppress warnings within these files (e.g., using specific zizmor: ignore comments).\n
Capability inventory: The skill allows the agent to execute command-line tools including zizmor, pinact, actionlint, and git to analyze and commit changes to the repository.\n
Sanitization: The skill includes specific rules to improve sanitization, such as moving expressions to environment variables to prevent shell injection, as documented in references/rule-template-injection.md.\n- [EXTERNAL_DOWNLOADS]: The documentation provides instructions for downloading external tools required for workflow linting.\n
Evidence: references/rule-unpinned-images.md contains a shell script template that fetches the actionlint binary from its official GitHub repository: https://github.com/rhysd/actionlint/releases/download/v${ACTIONLINT_VERSION}/actionlint_${ACTIONLINT_VERSION}_linux_amd64.tar.gz.\n
Context: The download process is secured by a hardcoded SHA256 checksum verification (sha256sum --check) before the binary is extracted or used.\n- [PRIVILEGE_ESCALATION]: The skill suggests environment setup scripts that require elevated permissions for package management.\n
Evidence: SKILL.md includes a bash snippet for bin/setup that utilizes sudo pacman -S --noconfirm to install dependencies on specific Linux distributions.\n
Context: This is a standard administrative task for developer machine setup and is intended for the primary purpose of tool installation.

harden-github-actions