edl-edit

Warn

Audited by Gen Agent Trust Hub on Nov 9, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructions direct the agent to execute multiple shell commands using the uv run video-agent prefix. These commands perform file operations, render video outputs, and run transcription tools.- [EXTERNAL_DOWNLOADS]: The skill depends on an external toolkit hosted at https://github.com/agamm/video-agent. The documentation explicitly instructs the agent to clone this repository to access the necessary scripts, creating a dependency on unverified third-party code.- [INDIRECT_PROMPT_INJECTION]: The skill processes external data by transcribing output videos and reviewing the resulting text. This workflow is vulnerable to malicious instructions embedded in video audio that could influence the agent's behavior during the verification phase.
  • Ingestion points: Processes rationales in edit.json and transcriptions of output video files (out.mp4).
  • Boundary markers: The skill does not define explicit delimiters to isolate external transcript content from agent instructions.
  • Capability inventory: Executes shell commands for rendering and transcription via uv run.
  • Sanitization: No sanitization or filtering of transcribed text is described before it is processed by the agent.- [DYNAMIC_EXECUTION]: The EDL schema allows raw FFmpeg filter chains to be defined in the audio_fix and grade fields. This provides a surface for potential command injection or unauthorized file access via complex FFmpeg filter logic if untrusted data is included in the configuration.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 9, 2026, 12:32 PM