gepetto
Pass
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from multiple external and internal sources to generate comprehensive implementation plans, creating a surface for indirect injection attacks.
- Ingestion points: The workflow ingests content from user-supplied markdown specifications, web search results (via the Explore subagent), local codebase analysis (via the Explore subagent), and feedback provided by external LLM reviews (Gemini and Codex).
- Boundary markers: The instructions do not define robust delimiters or explicit "ignore embedded instructions" warnings for the subagents when they are processing and synthesizing findings from web and codebase research.
- Capability inventory: The skill possesses the ability to write numerous markdown files, execute shell commands (for external reviews), and generate high-density prompts intended for autonomous execution by other agents such as ralph-loop or Ralphy.
- Sanitization: There is no evidence of explicit sanitization, filtering, or validation of content retrieved from the web or codebase before it is integrated into the final 'claude-plan.md' and its corresponding implementation sections.
- [COMMAND_EXECUTION]: The skill uses shell commands to interact with external model review tools during the planning phase.
- Evidence: The 'external-review.md' protocol instructs the agent to execute the
geminiandcodexCLI tools via bash, passing synthesized plan content through shell interpolation or stdin. While these tools are from trusted providers, the execution logic relies on the integrity of the plan content generated in previous steps. - [EXTERNAL_DOWNLOADS]: The skill documentation and logic reference external tools and repositories from well-known technology organizations.
- Evidence: The skill references and utilizes the Gemini CLI (github.com/google-gemini/gemini-cli) and OpenAI's Codex for plan validation. It also integrates with ralph-loop (an official Anthropic plugin) and Ralphy (an external CLI tool). These references target established services and organizations.
Audit Metadata