seedance-ad-creative
Pass
Audited by Gen Agent Trust Hub on Sep 22, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill interacts with the Ofox API at api.ofox.ai to submit video generation jobs, check pricing, and retrieve model data. These requests are routed through a dedicated utility script provided by the vendor.
- [COMMAND_EXECUTION]: Shell commands including bash, curl, and jq are used to facilitate API interactions and process data. The skill implements a path discovery loop to locate its required local dependency (ofox-video-core) in standard installation directories.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests user-supplied text and images to construct detailed prompts for a generative video model. While this creates a potential surface for indirect injection, the skill mitigates risk through the use of highly structured prompt manifests and templates that provide clear context boundaries for the model.
- Ingestion points: Untrusted data enters the agent context through product descriptions and image file paths provided by the user.
- Boundary markers: Prompts are organized using specific headers (e.g., STYLE, PRODUCT, SCENE, CAMERA) and a timeline manifest to isolate different instructions.
- Capability inventory: The skill reads local product images and uses network tools to communicate with the vendor's API.
- Sanitization: Input data is placed within a strictly defined prompt structure designed for creative steering rather than direct command processing.
Audit Metadata