The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches the LlamaGuard model weights from Meta's official HuggingFace repository (meta-llama/LlamaGuard-7b). This is an expected operation for using gated models provided by a well-known service.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user and assistant messages to perform safety classification.
Ingestion points: Untrusted data enters the agent context through the user_message and bot_response parameters in the moderate, check_input, check_output, moderate_vllm, and moderate_endpoint functions within SKILL.md.
Boundary markers: The skill uses tokenizer.apply_chat_template, which applies standard model-specific formatting to delimit roles and content, providing structural protection.
Capability inventory: No dangerous operations such as shell execution (subprocess), arbitrary code evaluation (exec/eval), or unexpected file system writes are performed based on the processed data; the output is used for safe conditional logic (blocking vs. allowing).
Sanitization: The skill relies on the LlamaGuard model's safety alignment to classify and mitigate malicious instructions embedded within the processed content.

llamaguard