The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted code from a user-specified scope to perform security audits. A malicious codebase could theoretically include instructions aimed at influencing the audit report, remediation steps, or the content of generated issues.
Ingestion points: The skill reads the full codebase or specific directories defined by the user in Step 1.
Boundary markers: No specific delimiters or "ignore instructions" warnings are defined for the audited content.
Capability inventory: The skill can write code changes (Step 8), execute shell commands for testing (Step 8), and interact with external APIs via the GitHub CLI (Step 9).
Sanitization: The skill does not explicitly describe sanitization or escaping of the code evidence before it is included in the audit report or GitHub issues.
[DYNAMIC_EXECUTION]: The skill uses a Python one-liner to parse JSON output from a shell command during the issue creation process.
Evidence: python3 -c "import json,sys; print(json.load(sys.stdin)['id'])" is used in Step 9 to extract a project item ID.
[COMMAND_EXECUTION]: The skill relies on the GitHub CLI (gh) to manage the lifecycle of security findings.
Evidence: Commands such as gh issue create, gh project item-add, gh project item-edit, and gh issue close are invoked in Steps 9 and 10.
[DATA_EXFILTRATION]: Audit findings, including code snippets and evidence of vulnerabilities, are transmitted to GitHub.
Evidence: Step 9 extracts evidence from the codebase and posts it to GitHub Issues and Projects. Since this targets a well-known service (GitHub) as part of a standard development workflow, this is documented as expected functionality.

security-audit