higgsfield-motion-design
Pass
Audited by Gen Agent Trust Hub on Sep 28, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided text such as brand names, taglines, and stylistic moods and interpolates them into prompt templates for image and video generation tools. This represents a standard functional surface for indirect prompt injection if the ingested data contains adversarial instructions.
- Ingestion points: User responses provided during the 'Brief intake' (Step 1), including mood, brand name, and tagline.
- Boundary markers: The prompt skeletons in Step 3 and Step 4 lack explicit delimiters or instructions for the model to ignore potential commands embedded within user-provided strings.
- Capability inventory: Access to
generate_image(GPT Image 2) andgenerate_video(Seedance 2.0) tools which generate AI media based on user input. - Sanitization: The skill does not define specific sanitization, validation, or escaping logic for the brand and mood strings before they are inserted into tool prompts.
Audit Metadata