The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it instructs the agent to ingest and process untrusted user data (text prompts and image URLs) to generate 3D assets.
Ingestion points: User input enters the skill's workflow through the prompt and file parameters in the text_to_model, image_to_model, and multiview_to_model tasks defined in SKILL.md.
Boundary markers: The skill lacks explicit instructions for delimiting or ignoring potentially malicious instructions embedded in the user-supplied prompts, although it does note character set limitations (no emoji/exotic Unicode).
Capability inventory: The skill utilizes network communication (curl commands) and file retrieval (downloading generated meshes from time-limited URLs) as described in SKILL.md.
Sanitization: No input validation or output sanitization routines are described in the provided instructions to mitigate adversarial content in user inputs.
[DATA_EXFILTRATION]: The skill performs network operations to external service domains to interact with the Tripo API.
Evidence: The skill documentation (SKILL.md) provides endpoints for https://api.tripo3d.ai/v2/openapi and instructs the agent on how to manage API keys at platform.tripo3d.ai.
Context: These network operations are necessary for the primary function of the skill (AI model generation) and target the official vendor infrastructure.

tripo-3d