mage-video-face-swap
Pass
Audited by Gen Agent Trust Hub on Oct 17, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses
ffprobeto measure the duration of video files (ffprobe -v error -show_entries format=duration -of csv=p=0 clip.mp4). This is a standard, non-malicious use of a command-line utility for processing media metadata. - [EXTERNAL_DOWNLOADS]: The skill interacts with
https://mcp.mage.space/mcpandhttps://www.mage.space. These are the official domains of the skill's author (mage-space) and are used for API communication and account management. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of video and image URLs provided by the user.
- Ingestion points: Video URLs, image URLs, and saved character handles enter the agent's context through user prompts.
- Boundary markers: The instructions use specific JSON schemas and a "wrapped prompt" template to isolate the user-provided inputs from the core instructions sent to the generation model.
- Capability inventory: Capabilities include
generate(API call to Mage),create_upload(file upload), andffprobe(local command execution for metadata). - Sanitization: The skill provides strict templates for the AI to follow ("Fill in only the marked slots; don't rewrite, translate, or add to it") to ensure user inputs do not override the system instructions.
Audit Metadata