carousels
Pass
Audited by Gen Agent Trust Hub on Nov 9, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands to process video and render slides.
- Evidence includes the use of
ffmpegto extract frames from video files (ffmpeg -ss ... -i "<finished-edit.mp4>" ...). - Evidence includes execution of the skill's own Python script to generate the final assets (
python3 render_carousel.py <config.json> <out_dir>). - [EXTERNAL_DOWNLOADS]: The skill requires external software and libraries to function.
- The documentation specifies that it needs
ffmpeginstalled on the system. - It directs the user or agent to install Python packages from public registries:
Pillow(required) andpilmoji(optional). - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, which presents a surface for indirect prompt injection.
- Ingestion points: The skill ingests transcripts from processed videos, content from YouTube URLs, and existing markdown files (
repurpose-plan.md,/content-cascadeposts) to determine carousel content. - Boundary markers: The instructions do not define strict boundary markers or sanitization for the ingested text before it is used to draft carousel slides or image generation prompts.
- Capability inventory: The agent has the capability to write files (PNG, PDF, JSON), execute local shell commands (ffmpeg, python), and invoke external image generation tools (
gpt_image_2). - Sanitization: No specific sanitization or filtering of the ingested content is defined in the instructions, though the output is limited to text rendering within images and document generation.
Audit Metadata