azure-ai-voicelive-dotnet
Pass
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill implements a real-time interaction workflow that ingests external data from both user messages and AI model responses (audio and text deltas).
- Ingestion points: Data enters the agent context via
session.GetUpdatesAsync()which processes server events, andsession.AddItemAsync()which accepts user inputs. - Boundary markers: The documentation includes a dedicated 'Voice and tool boundaries' section that explicitly warns the user to validate tool arguments, allowlist operations, and require authorization for external effects.
- Capability inventory: The skill utilizes network operations (WebSockets to Azure AI services) and local audio processing via the NAudio library.
- Sanitization: While the code snippets do not implement full sanitization, the instructions provide clear guidance on the necessity of validating untrusted model output before execution.
- [EXTERNAL_DOWNLOADS]: The skill references and provides installation instructions for external libraries and documentation.
- Evidence: The skill suggests installing
Azure.AI.VoiceLive,Azure.Identity, andNAudiovia NuGet. - Context: All referenced URLs and source repositories belong to Microsoft, Azure, or well-known package registries (NuGet). These are recognized as trusted organizations, and the references are documented neutrally as standard development practices.
Audit Metadata