skill-creator
Pass
Audited by Gen Agent Trust Hub on Aug 24, 2026
Risk Level: SAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill uses subprocess.Popen in scripts/run_eval.py to execute the claude CLI.
- Evidence: Found in scripts/run_eval.py where subprocess.Popen is called with the claude command and user-provided queries.
- [DYNAMIC_CONTEXT_INJECTION]: The skill documents the use of !command for shell output injection.
- Evidence: SKILL.md contains a section on 'Dynamic shell injection' explaining how to use !command to inject live shell results.
- [DYNAMIC_EXECUTION]: The skill creates temporary metadata files.
- Evidence: scripts/run_eval.py writes temporary .md skill files to .claude/commands/ during evaluation.
- [INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection through its ingestion of test cases and rendering of outputs.
- Evidence Chain: 1. Ingestion points: evals/evals.json and agent outputs. 2. Boundary markers: Not explicitly enforced. 3. Capability inventory: Shell execution via claude CLI. 4. Sanitization: html.escape used in viewer.html and generate_report.py.
Audit Metadata