The Agent Skills Directory

[SAFE]: The skill provides comprehensive instructional content and boilerplate code for utilizing LlamaGuard models for AI safety and alignment tasks.
[EXTERNAL_DOWNLOADS]: The skill includes instructions to download pre-trained model weights from official Meta repositories on Hugging Face. These downloads originate from a well-known service and a trusted organization.
[INDIRECT_PROMPT_INJECTION]: The skill contains an inherent attack surface for indirect prompt injection as its primary function is to process and classify untrusted user-supplied text.
Ingestion points: User-provided text is passed to the moderate, moderate_vllm, and moderate_endpoint functions via the chat or messages variables.
Boundary markers: The skill follows best practices by using tokenizer.apply_chat_template to format inputs according to the model's expected safety prompt structure.
Capability inventory: The skill requires network access for initial model downloading and utilizes local compute resources (GPU/CPU) for inference. It also includes an example of a FastAPI server exposing a network endpoint.
Sanitization: No explicit sanitization of input text is provided, as the model itself functions as the primary safety and filtering mechanism for the agent's pipeline.

llamaguard