skill-creator-advanced

Pass

Audited by Gen Agent Trust Hub on Mar 29, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTIONREMOTE_CODE_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The scripts scripts/run_eval.py and scripts/improve_description.py utilize Python's subprocess module to call the claude CLI. This is a core feature used to programmatically trigger evaluations and optimize skill descriptions based on LLM feedback. The scripts correctly handle process termination and output capture.\n- [REMOTE_CODE_EXECUTION]: The scripts/run_eval.py script executes test queries loaded from an external JSON evaluation set by passing them to the subagent via the claude CLI. While this represents the execution of instructions from an external source, it is restricted to the local CLI environment and is the intended method for benchmarking triggering behavior.\n- [PROMPT_INJECTION]: The skill implements iterative loops that ingest data from local project files and evaluation results into prompts for a subagent. This constitutes an indirect prompt injection surface.\n
  • Ingestion points: Evaluation queries from JSON files and instructions/resources from SKILL.md and related metadata.\n
  • Boundary markers: scripts/improve_description.py uses XML-like tags to delimit ingested content, though run_eval.py passes queries directly to the CLI command.\n
  • Capability inventory: Subprocess execution of the claude CLI for task performance and evaluation.\n
  • Sanitization: Candidate descriptions are normalized in scripts/improve_description.py to remove angle brackets to prevent tag injection or malformed XML generation.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 29, 2026, 05:29 AM