The Agent Skills Directory

[PROMPT_INJECTION]: The reference files contain examples of prompt injection payloads and instructions on how to defend against them. For example, llm-security.md describes "Indirect prompt injection" to teach the auditor how to spot it. These are documentation artifacts used for educational and auditing purposes, not actual attacks against the agent.
[COMMAND_EXECUTION]: The skill includes examples of dangerous commands (e.g., rm -rf /, chmod 777) within its educational references to demonstrate vulnerabilities like Command Injection (CWE-78) and destructive behavior. These strings are within documentation text and are not executed by the skill itself.
[DATA_EXFILTRATION]: While the skill uses the GitHub CLI (gh) and provides patterns for using trufflehog (which can verify secrets over the network), these actions are strictly scoped to the skill's purpose of security auditing. The secrets.sh script explicitly checks for a specific environment variable before enabling network-based secret verification, defaulting to local-only scanning. No unauthorized exfiltration to external domains was detected.
[EXTERNAL_DOWNLOADS]: Documentation in automated-scanning.md provides a secure example of downloading the opengrep binary from its official GitHub repository, including SHA256 checksum verification. These downloads are from a trusted organization and are provided as part of a secure implementation guide for the user.

security-audit