metaclaw-setup-architect

Warn

Audited by Gen Agent Trust Hub on Mar 3, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCREDENTIALS_UNSAFE
Full Analysis
  • [COMMAND_EXECUTION]: The skill is designed to generate BOOTSTRAP.md files for agents, which are explicitly instructed to include 'actual shell commands' for system configuration and dependency installation. These scripts are intended to execute on the first run, creating a direct path for command execution.
  • [REMOTE_CODE_EXECUTION]: The skill facilitates the generation of openclaw.json configurations that define MCP (Model Context Protocol) servers. These definitions often use npx or uvx to download and execute remote packages (e.g., @anthropic/mcp-server-brave-search).
  • [CREDENTIALS_UNSAFE]: The architecture manages sensitive credentials such as DISCORD_BOT_TOKEN, TELEGRAM_BOT_TOKEN, and OWNER_PHONE. While the templates correctly advocate for referencing environment variables (e.g., env:TOKEN), the `Phase 4
  • Installprocess instructs the agent to provide explicitexport` commands, which may lead to secrets being stored in shell history files.
  • [EXTERNAL_DOWNLOADS]: The generated setup instructions and skill templates include commands to download various third-party packages and binaries using package managers like pip, npm, and brew (e.g., yt-dlp, openai-whisper, ffmpeg).
  • [PROMPT_INJECTION]: As a meta-generator, the skill is susceptible to indirect prompt injection. Malicious data provided during the 'Discovery' phase or ingested by the generated agents from external sources (like scraped web content or YouTube transcripts) could influence the generated logic, instructions, or shell commands.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 3, 2026, 06:57 AM