poetry-cinema-page

Warn

Audited by Gen Agent Trust Hub on Sep 22, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: Multiple scripts within the skill utilize the subprocess.run function to execute local command-line tools. scripts/video.py drives the production pipeline by calling node, npm, and remotion for video rendering, while both scripts/video.py and scripts/narration.py execute ffmpeg and ffprobe for audio/video processing tasks.
  • [EXTERNAL_DOWNLOADS]: The skill performs several types of external network operations. scripts/video.py runs npm install to download dependencies from the public registry. scripts/ark_image.py interacts with Volcengine Ark and an unofficial third-party gateway (image.mlgb7.com) to generate and download image assets at runtime.
  • [REMOTE_CODE_EXECUTION]: The execution of npm install in scripts/video.py represents a significant risk as it allows for arbitrary code execution on the local system through package lifecycle scripts (e.g., preinstall, postinstall) during the dependency installation phase.
  • [DATA_EXFILTRATION]: The skill is designed to send user-supplied API keys (such as GPT_IMAGE_API_KEY) to an unofficial third-party relay service located at image.mlgb7.com. Sending sensitive credentials to an untrusted domain increases the risk of credential theft or unauthorized access if the relay is compromised.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted poetry text provided by the user, creating a surface for potential injection attacks.
  • Ingestion points: Untrusted poetry text is ingested in scripts/ark_image.py and scripts/narration.py from files like narration.txt or via direct command-line arguments.
  • Boundary markers: Absent; the skill does not utilize delimiters or specific instructions to isolate untrusted input from the surrounding prompt templates.
  • Capability inventory: The skill has extensive local capabilities including command execution (shell access via subprocess), file system modifications, and network access to multiple external domains.
  • Sanitization: No validation or sanitization is performed on the ingested text before it is interpolated into prompts or command-line parameters.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 22, 2026, 05:56 PM