vlog-auto-edit
Pass
Audited by Gen Agent Trust Hub on May 1, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill provides Python snippets and instructions that use subprocess.run with shell=True to execute FFmpeg commands. This pattern appears in SKILL.md for clip trimming and highlight montage generation. While these are part of the intended functionality, executing shell commands with interpolated filenames (e.g., f'ffmpeg ... "{f}"') can be risky if filenames are not properly sanitized.- [PROMPT_INJECTION]: The workflow is vulnerable to indirect prompt injection (Category 8).
- Ingestion points: Audio from user-provided video files is transcribed using ASR engines (Whisper or FunASR) and stored in clip_analysis.json.
- Boundary markers: The transcript text is placed inside JSON structures and subsequently inside Markdown code blocks within the edit_plan_prompt.md template, but there are no specific instructions for the LLM to ignore or treat embedded transcript text as untrusted data.
- Capability inventory: The agent has the ability to execute shell commands (ffmpeg), write files, and interact with network APIs.
- Sanitization: There is no evidence of sanitization or filtering applied to the transcribed text before it is interpolated into the prompt for the narrative arrangement stage.- [EXTERNAL_DOWNLOADS]: The skill involves several remote resource interactions. It installs third-party Python packages from standard registries and downloads large pre-trained models (~1GB) from GitHub (OpenAI Whisper) and ModelScope (Alibaba) upon first use. Additionally, it sends base64-encoded image frames from the user's footage to external vision APIs (e.g., Zhipu AI, OpenAI, Alibaba) for scene understanding.
Audit Metadata