The Agent Skills Directory

[SAFE]: The skill provides documentation and implementation examples for LlamaGuard, a specialized content moderation model from Meta. All external resources, including model weights accessed via HuggingFace and dependencies like transformers, torch, and vllm, are sourced from well-known technology organizations and official registries.
[INDIRECT_PROMPT_INJECTION]: Untrusted user data is ingested through the moderate, check_input, and check_output functions defined in SKILL.md. The implementation utilizes the tokenizer.apply_chat_template method to establish structural boundary markers for this content. An inventory of the code confirms a lack of dangerous capabilities such as subprocess calls, arbitrary execution (eval/exec), or unauthorized network exfiltration. The skill is specifically designed to identify and filter malicious content, acting as a defensive layer against prompt injection attacks.

llamaguard