The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text and media inputs for asset generation, which introduces a surface for data-driven prompt injection.
Ingestion points: User instructions for texture generation (texture_prompt in references/3d.md), audio synthesis (instruction in references/audio.md), and general user queries for video analysis are ingested by the skill.
Boundary markers: There are no explicit delimiters or specific 'ignore-embedded-instructions' markers defined in the referenced files to isolate user-supplied prompts from the model's task instructions.
Capability inventory: The skill utilizes tools for dynamic model exploration (models_explore), job status monitoring (job_status), and remote media generation, which are triggered based on these user inputs.
Sanitization: The skill effectively mitigates direct file injection by explicitly rejecting raw external URLs and requiring internal media_id or job_id references for all input media assets.

media-higgsfield-assets