The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains various strings used for red-teaming and evaluation purposes.
Evidence: Found in agents/test-writer.md and references/testing.md. These include examples like 'Ignore previous instructions' and 'Repeat the system prompt'. These are explicitly used to verify agent robustness and prevent prompt leaking in the generated test suites.
[EXTERNAL_DOWNLOADS]: The skill references a URL that is flagged as blacklisted by security scanners.
Evidence: Found in agents/auth-upgrader.md. The URL https://auth.your-server.com is used as a placeholder in example OAuth metadata configuration. Referencing blacklisted domains, even as placeholders, carries a risk of unintended network traffic or user confusion.
[COMMAND_EXECUTION]: The skill utilizes high-risk tools and generates instructions for complex system operations.
Evidence: SKILL.md requires the Bash tool for project detection and environment verification. agents/context-writer.md generates logic to manage processes, including commands like xargs kill for troubleshooting. This is consistent with the skill's purpose of managing developer infrastructure.
[DATA_EXFILTRATION]: The skill processes untrusted codebase data and has the capability to write files and execute network requests.
Evidence: The 'Audit Workflow' involves reading project files via Glob and Grep to score agent readiness. This ingestion of untrusted local data combined with the ability to dispatch sub-agents and execute shell commands creates a surface for indirect prompt injection if the audited codebase contains malicious instructions targeting the auditor.

surface