ad-creative
Pass
Audited by Gen Agent Trust Hub on Nov 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data verbatim to generate ad creative, creating a surface for potential injection.
- Ingestion points: Reads customer reviews from
inputs/reviews/and campaign comments frominputs/comments/to ground copy in authentic language (SKILL.md). - Boundary markers: The
creative-review-page.mdreference andcreative-review-template.htmlinclude specific instructions to escape HTML characters (e.g., using\u003cand a customesc()function) when populating the review artifact to prevent script injection. - Capability inventory: The skill performs file writes to dated output folders, executes network requests via
curlto external AI APIs, and manages ad platform campaigns using internalnodeCLI tools (SKILL.md, references/generative-tools.md). - Sanitization: Implements an
escfunction in the HTML template's JavaScript to sanitize data before rendering to the DOM. - [EXTERNAL_DOWNLOADS]: The skill references and provides instructions for utilizing several external tools, libraries, and services.
- Includes setup instructions for
npx gooseworks install,npx create-video(Remotion), and cloning thevoiceboxrepository from GitHub (references/imessage-video-ads.md, references/generative-tools.md). - Mentions numerous well-known AI services (Google Gemini, OpenAI, ElevenLabs, Runway, Flux, Ideogram) for generating ad assets.
- [COMMAND_EXECUTION]: Multiple shell command examples are provided for the agent or operator to perform administrative and generative tasks.
- Provides
curlexamples for interacting with external APIs andnodecommand examples for campaign management via provided CLI tools (SKILL.md, references/generative-tools.md). - [DYNAMIC_EXECUTION]: The skill generates a self-contained HTML review page that executes JavaScript to render ad mockups.
- The
assets/creative-review-template.htmlfile uses a data-driven rendering approach where concept data is parsed from a JSON block at runtime. The skill specifically instructs the agent on how to populate this data safely to avoid breaking the execution context.
Audit Metadata