ad-creative
Pass
Audited by Gen Agent Trust Hub on Aug 5, 2026
Risk Level: SAFEREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill's reference guide (
references/generative-tools.md) documents installation steps for 'Voicebox' by cloning from an external GitHub repository and running build scripts (make setup). While presented as documentation for the user, an automated agent might attempt these commands if requested to set up the environment.\n- [COMMAND_EXECUTION]: The skill provides examples of executing local CLI tools (e.g.,node tools/clis/google-ads.js) to fetch campaign data and manage ad performance. These commands assume a pre-configured environment with these scripts available.\n- [EXTERNAL_DOWNLOADS]: The skill references and provide code snippets for numerous well-known AI services including Google Gemini, Replicate, ElevenLabs, and Runway. These integrations use standard REST API patterns with environment-based credential management (e.g.,$GEMINI_API_KEY).\n- [PROMPT_INJECTION]: The skill processes external performance data (CSV, API output, or text) to generate creative iterations. This creates an indirect prompt injection surface.\n - Ingestion points: Performance data provided via CSV, API, or paste (
SKILL.md).\n - Boundary markers: None; the skill lacks specific delimiters or instructions to disregard commands embedded in the performance data.\n
- Capability inventory: The agent uses ad platform CLI tools to read performance reports and generate creative content.\n
- Sanitization: No evidence of input validation or escaping for the ingested performance data.
Audit Metadata