The Agent Skills Directory

[SAFE]: The skill functions as a documentation resource providing security architecture and best practices. It does not contain any executable malicious code or instructions intended to compromise the host environment.
[PROMPT_INJECTION]: The skill mentions prompt injection patterns such as "ignore previous instructions" and "system prompt". These are used exclusively within educational examples to teach security practitioners how to detect and mitigate injection attacks using validation logic and regex patterns. They do not constitute an active injection attempt against the agent.
[EXTERNAL_DOWNLOADS]: The skill references trusted and well-known components, such as official GitHub Actions (e.g., actions/checkout, step-security/harden-runner) and official Model Context Protocol (MCP) packages. These references are standard for the described security workflows.
[COMMAND_EXECUTION]: Technical examples include shell scripts for pre-commit hooks and configuration for GitHub Actions runners. These are provided as templates for defense-in-depth and sandboxing, not for malicious execution.
[DATA_EXFILTRATION]: The skill discusses sensitive file paths (e.g., .env, ~/.ssh/, /etc/passwd) and secret detection regexes. These references are used to demonstrate how to configure access control lists (ACLs) and Data Loss Prevention (DLP) scanners to prevent unauthorized data exposure.
[DYNAMIC_EXECUTION]: Examples of dynamic code execution functions like eval() and new Function() are provided specifically to warn developers against their use and to show how static analysis tools can detect them.

GitHub Agentic Workflows Security Architecture