The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill architecture converts audio input to text and passes it to a Large Language Model (LLM) for processing. This creates a vulnerability where spoken instructions in the audio stream could maliciously influence the AI's behavior.
Ingestion points: VoiceAssistant.process_voice_input in examples/voice_assistant.py transcribes audio files or streams into text strings.
Boundary markers: The skill lacks delimiters or "ignore instructions" warnings when interpolating the transcribed user speech into the conversation_history used for LLM response generation.
Capability inventory: The skill manages conversation state and interacts with multiple external APIs, which could be exploited through successful prompt injection.
Sanitization: No filtering or validation is performed on the transcribed text before it is processed by the AI.
[EXTERNAL_DOWNLOADS]: The skill communicates with external APIs to process audio data, including AssemblyAI (api.assemblyai.com) for transcription and ElevenLabs (api.elevenlabs.io) for speech synthesis. These are established services for voice-related tasks.

voice-ai-integration