higgsfield-faceless-video
Pass
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill relies on executing local scripts and system utilities such as ffmpeg, ffprobe, and the higgsfield CLI to process video and audio components. This execution is fundamental to the skill's primary purpose of media assembly.\n- [EXTERNAL_DOWNLOADS]: The assemble_blocks.sh script contains a fetch function that uses curl to retrieve external assets like background music or video clips from remote URLs. These assets are then used as inputs for the assembly process.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process user-supplied content, such as scripts, lyrics, and topics, which are interpolated into prompts for generative models. While the skill employs significant validation logic through scripts like validate_motion_script.py and validate_song_prompt.py to ensure adherence to formatting and timing constraints, it possesses an inherent surface for indirect prompt injection.\n
- Ingestion points: User-provided manifests including script.json, story.json, and song_prompt.txt (referenced in references/channel-motion.md and references/song.md).\n
- Boundary markers: The instructions do not specify explicit boundary markers or delimiters when passing these user-provided strings to generative models.\n
- Capability inventory: The skill has capabilities for file system writes, network fetches (curl), and subprocess execution (ffmpeg, python3).\n
- Sanitization: The skill uses complex validation scripts to verify that inputs meet specific word count, duration, and linguistic criteria before generation begins.
Audit Metadata