paper-collage-ad

Pass

Audited by Gen Agent Trust Hub on Jul 10, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill downloads the 'mlx-indextts' repository from GitHub and pre-trained voice models from HuggingFace. These are utilized locally on the user's machine for authorized voice cloning and narration generation.
  • [COMMAND_EXECUTION]: The automation pipeline extensively uses system commands including 'ffmpeg' and 'ffprobe' for media rendering, 'node' for manifest processing, and 'bash' for dependency management and environment setup.
  • [REMOTE_CODE_EXECUTION]: The skill employs package runners like 'uvx' and 'npx' to execute remote tools such as 'demucs' (for audio track separation) and 'hyperframes' (for rendering HTML-based animations). These are functional dependencies for the advertisement production workflow.
  • [DATA_EXFILTRATION]: Functional data transmission is used to communicate with AI services. The skill sends prompts, text, and images to established providers including Volcengine (Doubao/Jimeng), ElevenLabs, and MiniMax to generate project assets. Access keys for these services are managed via local environment variables.
  • [PROMPT_INJECTION]: The skill processes untrusted product briefs and brand materials to generate video content, which represents an Indirect Prompt Injection surface. To mitigate this, the workflow enforces a 'HARD GATE' requiring human approval of the final script before any automated generation or command execution occurs.
  • [CREDENTIALS_UNSAFE]: No hardcoded credentials were found. The skill includes a dedicated 'privacy-check.sh' script and a project-level '.gitignore' template to help users avoid tracking API keys, private voice models, or sensitive brand assets in public repositories.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 10, 2026, 06:54 AM
Security Audit — agent-trust-hub — paper-collage-ad