stitch-loop
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses the
Bashtool to start a local development server usingnpx serve site/publicto visually verify generated pages. - [EXTERNAL_DOWNLOADS]: The skill downloads HTML and image assets from remote URLs (
htmlCode.downloadUrlandscreenshot.downloadUrl) provided by the Stitch MCP tool at runtime. These files are subsequently integrated into the project's public directory. - [INDIRECT_PROMPT_INJECTION]: The skill contains an indirect prompt injection surface because it ingests instructions and context from project files without sufficient sanitization or boundary markers.
- Ingestion points: The agent reads the baton file (
.stitch/next-prompt.md) for task instructions and consults project state in.stitch/SITE.mdand.stitch/DESIGN.md(SKILL.md, Steps 1 and 2). - Boundary markers: None identified; the instructions treat the content of these files as authoritative commands for the agent's next action.
- Capability inventory: The skill utilizes
Bashfor running servers,Writefor file system modifications, andstitch*:*tools for generating remote content. - Sanitization: The skill does not describe any validation or filtering of the content read from these files before using it to construct prompts for the generation tools.
Audit Metadata