creating-portable-skills

Warn

Audited by Gen Agent Trust Hub on Sep 4, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill workflow recommends downloading a Python-based validator from a pinned commit in the agentskills/agentskills GitHub repository if the tool is not already available on the system.
  • [COMMAND_EXECUTION]: The skill requires the execution of the skills-ref validate tool and performs installation checks ('smoke checks') that involve running shell commands across different agent harnesses like Claude Code, Codex, and Gemini CLI.
  • [INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes untrusted user input, history, and external documentation to generate new instructions and scripts. It possesses the capability to write files and execute commands. Mitigation steps include defined System-Owned Invariants and a requirement for independent fresh-context reviews.
  • Ingestion points: SKILL.md processes conversation history, user corrections, examples, and documentation.
  • Boundary markers: Employs System-Owned Invariants and independent review mandates in references/review-checklist.md.
  • Capability inventory: Performs file-system writes (Step 3) and command execution (Steps 4 and 8) in SKILL.md.
  • Sanitization: Uses structural validation, binary grading of behavioral cases, and fresh-context reviews to verify generated outputs.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 4, 2026, 08:31 PM
Security Audit — agent-trust-hub — creating-portable-skills