The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill demonstrates patterns for building voice assistants that ingest untrusted user input (transcribed speech) and use it to perform RAG lookups or generate replies, which is a standard surface for indirect prompt injection.
Ingestion points: The references/workflows.md file shows a lifecycle hook on_user_turn_completed that ingests new_message.text_content() directly from the user participant.
Boundary markers: The provided code examples do not demonstrate the use of delimiters or 'ignore' instructions when interpolating user content into prompts.
Capability inventory: The skill environment provides capabilities for speech generation (session.say), reply generation (session.generate_reply), and the execution of arbitrary Python functions via @function_tool.
Sanitization: There is no evidence of transcription sanitization, filtering, or validation before the data is processed by the LLM or used in database lookups.

agents-py