The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill includes instructions to download the official GitHub Actions runner package from GitHub's releases for setting up self-hosted runners.
[REMOTE_CODE_EXECUTION]: The skill provides instructions for downloading and executing shell scripts associated with the GitHub Actions runner. These resources originate from a well-known, trusted service provider.
[COMMAND_EXECUTION]: The skill demonstrates the standard use of the run keyword to execute build, test, and deployment scripts within CI/CD pipelines.
[INDIRECT_PROMPT_INJECTION]: The skill illustrates how to create workflows that process external inputs (e.g., via workflow_dispatch). These workflows have high capabilities, including shell execution and access to repository secrets.
Ingestion points: Workflow dispatch inputs and environment variables.
Boundary markers: No specific delimiting markers are shown in the examples to isolate external data from instructions.
Capability inventory: Workflows can execute shell commands, perform network operations via actions, and access credentials in ${{ secrets }}.
Sanitization: Explicit input sanitization or validation is not demonstrated.
[DATA_EXFILTRATION]: The debugging section suggests printing the full github and job JSON contexts to workflow logs. Users should be cautious as this could expose non-secret sensitive metadata in the build execution history.

github-actions