The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill installs official vendor tools and SDKs, including the Novita AI Sandbox SDK (novita-sandbox) via pip and the CLI tool (novita-sandbox-cli) via npm. These are vendor-provided resources necessary for the skill's functionality. It also references standard observability and framework libraries like langchain-openai, litellm, and py3nvml from official package registries.
[COMMAND_EXECUTION]: The skill uses curl to interact with the Novita AI API endpoints for model discovery, task submission, and polling. These operations are directed at the vendor's official domain (api.novita.ai) and are required for the service to function.
[DYNAMIC_EXECUTION]: The skill integrates the Novita AI Sandbox, which allows for isolated execution of code and file operations. This is an intended feature used for building autonomous agents in a secure, containerized environment, separated from the host system.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection as it processes user-provided text, images, and audio for interaction with Large Language Models.
Ingestion points: User-supplied prompts, local or remote media files (audio, video, images) passed to multimodal APIs.
Boundary markers: Explicit instructions in SKILL.md to use trusted local sources and a security warning in the LLM reference against embedding unverified content in system prompts serve as architectural guardrails.
Capability inventory: The skill possesses capabilities for network access via curl, file system reading, and code execution within the isolated Sandbox environment.
Sanitization: Documentation specifically recommends enabling NSFW detection for applications processing user-generated content.

novita-ai