stitch-design-md
Pass
Audited by Gen Agent Trust Hub on Oct 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external data (HTML, CSS, and screenshots) to generate design system documentation, which creates a potential surface for instructions embedded in data to influence the agent.
- Ingestion points: The skill processes local HTML files, raw code inputs, and remote assets fetched via web_fetch using URLs provided by the Stitch MCP.
- Boundary markers: The instructions mandate a specific six-section document structure and require the agent to explicitly label unverified inferences as NOT_VERIFIED.
- Capability inventory: Capabilities include network reading (web_fetch), interaction with the Stitch MCP tools, and writing documentation (DESIGN.md) to the local filesystem.
- Sanitization: The skill contains explicit rules for removing PII (customer names, phone numbers) and prohibits the inclusion of secrets, cookies, or signed URLs in the output.
- [EXTERNAL_DOWNLOADS]: The skill retrieves design assets and references documentation from established and trusted infrastructure.
- Evidence: Fetches project assets from the stitch.withgoogle.com domain and refers to the google-labs-code/stitch-skills repository for implementation logic.
Audit Metadata