The Agent Skills Directory

Indirect Prompt Injection (LOW): The skill implements a tool execution loop that processes untrusted data. \n
Ingestion points: user_message in scripts/function-def.py accepts external input. \n
Boundary markers: The script lacks explicit delimiters or instructions to isolate user input from the model's tool-calling logic. \n
Capability inventory: The registry.execute function allows the execution of registered Python functions based on LLM decisions. In the provided example, the tools (search_documents, get_weather) are read-only and have low impact. \n
Sanitization: The implementation uses OpenAI's strict mode and JSON schema validation to ensure structured arguments, which prevents schema confusion but does not validate the semantic safety of the input text.

function-calling