The Agent Skills Directory

Handling of Untrusted Data: The skill is designed to process real-time audio and text input from users and responses from AI models. This functionality establishes a surface area for indirect prompt injection, where malicious instructions embedded in user audio or model outputs could potentially influence the agent's behavior, particularly when using the integrated function-calling (tools) capabilities.
Ingestion points: Data enters the context via conn.input_audio_buffer.append and through server events received from the model connection.
Boundary markers: The documentation recommends using session instructions to establish agent behavior boundaries, though these do not fully eliminate injection risks.
Capability inventory: The skill facilitates network communication via WebSockets and supports the execution of external tools through a defined function schema.
Sanitization: The implementation relies on developers to provide specific validation and safety logic within their function call handlers.
Vendor-Specific External Dependencies: The skill involves the installation of official Azure and established libraries such as azure-ai-voicelive, azure-identity, and aiohttp. These packages are necessary for the skill's intended purpose and originate from recognized sources.
Network Communication with Cloud Services: The skill establishes connections to Azure Cognitive Services endpoints (e.g., api.cognitive.microsoft.com). This connectivity is essential for real-time AI processing and utilizes standard authentication protocols like DefaultAzureCredential as recommended by the vendor.

azure-ai-voicelive-py