auto-polish
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes Node.js scripts (
scan-arsenal.mjs,picker.mjs) to discover installed tools and serve a local interface. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted metadata from other local skills.
- Ingestion points:
scripts/scan-arsenal.mjsreadsSKILL.mdfrontmatter from all directories in~/.claude/skills/and server configurations from~/.claude/settings.json. - Boundary markers: Lacks explicit boundary markers or warnings when subagents process the extracted metadata.
- Capability inventory: The skill orchestrates multi-agent workflows and executes shell commands for building variants and running tools.
- Sanitization: Ingested skill descriptions and titles are used directly in prompts and UI without validation.
- [INDIRECT_PROMPT_INJECTION]: The local server in
scripts/picker.mjsenablesAccess-Control-Allow-Origin: *. A malicious website could potentially trigger the/saveendpoint to overwrite theselection.jsonconfiguration file, influencing the agent's subsequent tool selection. - [COMMAND_EXECUTION]: The skill scans sensitive local files like
~/.claude/settings.jsonand the environmentPATHto locate and inventory installed design tools.
Audit Metadata