The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill requires the installation of the following Python packages: azure-ai-voicelive, aiohttp, and azure-identity. The documentation files in the references/ directory are recovered from Microsoft's public skills repository, which is a trusted source.
[INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection as it processes external, untrusted data including real-time audio input, speech-to-text transcripts, and outputs from external tools (Function Tools and MCP).
Ingestion points: Untrusted data enters the agent context through the input_audio_buffer (audio streams) and the conversation.item.create method (which handles function call outputs and message content).
Boundary markers: The skill uses the instructions field within the RequestSession model to define system-level behavior and constraints.
Capability inventory: The skill possesses the capability to perform bidirectional network operations via WebSockets to Azure endpoints and execute local function calls based on model instructions.
Sanitization: The skill's documentation explicitly advises users to perform server-side validation of tool names, JSON schemas, and authorization, noting that generated function calls should be treated as requests rather than granted permissions.
[COMMAND_EXECUTION]: The SDK facilitates the execution of function tools. The provided examples show how to define and handle these calls locally within the user's application logic, emphasizing that function execution remains under the developer's control.
[DATA_EXFILTRATION]: As designed for a voice AI service, audio data and transcripts are transmitted to Azure Cognitive Services endpoints. The skill includes a 'Runtime and privacy boundaries' notice advising users to obtain proper authorization for recording and to minimize data retention.

azure-ai-voicelive-py