media-higgsfield-explainer
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill references the Higgsfield Skills public repository on GitHub (github.com/higgsfield-ai/skills) as a source for its logic, licensing (MIT), and prompt templates. This is a standard reference to a well-known service for documentation and baseline behavior.
- [INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from user queries and web investigation results (R-step) to generate narration and video prompts.
- Ingestion points: User-supplied topics, documents, and external web search results are ingested during the research phase and scriptwriting steps.
- Boundary markers: The skill uses rigid English templates (STYLE REFERENCE, SCENE, MOTION, AUDIO, NEGATIVE) to wrap generated content, which helps maintain control over the model's output, though it lacks explicit "ignore embedded instructions" warnings within the templates.
- Capability inventory: The skill utilizes specialized MCP tools for audio, image, and video generation (
generate_audio,generate_image,generate_video). - Sanitization: A critical human-in-the-loop gate (Phase 3.5) is implemented where the user must approve the full narration and prompt manifest before any generation tools are invoked, significantly mitigating the risk of executing malicious instructions hidden in input data.
Audit Metadata