The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data (subject, action) and external URLs, providing a surface for indirect prompt injection attacks.
Ingestion points: User-provided subjects, actions, and HTTPS URLs referenced in the run_model tool instructions within SKILL.md.
Boundary markers: The instructions do not specify the use of delimiters or clear markers to separate untrusted user content from agent instructions.
Capability inventory: The skill uses run_model, get_model_request_status, and get_model_request_result to process and retrieve content.
Sanitization: Instructions emphasize schema matching for technical parameters but do not explicitly mandate sanitization of natural language prompts before submission to the model.
[DATA_EXFILTRATION]: The skill permits network operations to external, non-whitelisted domains by requiring publicly accessible HTTPS URLs for processing.
Evidence: The instruction "Use publicly accessible HTTPS URLs for model file inputs" in SKILL.md enables the agent to interact with arbitrary external endpoints.

generate-video