skill-creator

Pass

Audited by Gen Agent Trust Hub on Sep 9, 2026

Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONPRIVILEGE_ESCALATIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The skill's scripts (run_eval.py, improve_description.py, generate_review.py) use subprocess to invoke the claude CLI and system tools like lsof. These tools are used for running test cases and managing the review server.
  • [DYNAMIC_EXECUTION]: run_eval.py generates temporary markdown files in the .claude/commands/ directory to simulate skill loading for triggering tests. This involves creating and then deleting executable-equivalent configurations.
  • [INDIRECT_PROMPT_INJECTION]: The skill acts as a generator for other agent skills, processing user tasks and feedback to derive instructions, which presents a surface for indirect prompt injection.
  • Ingestion points: Reads task prompts from evals.json and user reviews from feedback.json.
  • Boundary markers: The SKILL.md includes safety-oriented instructions to avoid generating harmful skills, though these are natural language directives.
  • Capability inventory: Includes file system modification, command execution, and subagent task spawning.
  • Sanitization: Inputs are generally handled as list elements for subprocess calls, though generated descriptions are embedded into markdown files during triggered testing.
  • [PRIVILEGE_ESCALATION]: The review viewer management script (generate_review.py) searches for and terminates processes on specific ports using lsof and os.kill. Additionally, run_eval.py explicitly removes the CLAUDECODE environment variable to bypass nesting restrictions in the claude CLI.
  • [EXTERNAL_DOWNLOADS]: The skill's evaluation viewer (viewer.html) fetches the SheetJS library from a public CDN to facilitate Excel file rendering within the browser UI.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 9, 2026, 10:05 AM