fal-video-production

Pass

Audited by Gen Agent Trust Hub on Nov 7, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the user brief and on-screen text fields, which are subsequently used to generate prompts for external media generation models. It also processes outputs from vision and transcription tools during the quality check stage. The skill includes multiple guardrails and boundary markers (e.g., 'Never invent on-screen text', 'Never pad the script to fill the runtime') to mitigate the risk of the agent following instructions embedded in the source data. The capabilities associated with this surface include network calls to generation APIs and media processing via command-line tools.
  • [COMMAND_EXECUTION]: The skill executes shell-based commands for ffmpeg and utilizes the PIL (Pillow) library within a sandbox environment to perform media assembly tasks such as concatenation, audio sidechain compression (ducking), and burning text overlays onto video frames. These operations are performed on local stems and temporary artifacts generated during the production pipeline.
  • [EXTERNAL_DOWNLOADS]: The workflow involves extensive coordination with remote AI endpoints provided by the vendor (e.g., fal-video-generation, fal-audio-generation, fal-ai/video-understanding). These connections are used to fetch generated media assets and analysis results, which is the primary purpose of the skill.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 7, 2026, 07:17 AM