The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides templates for building voice agents that process untrusted audio input from users, creating a vulnerability surface where transcribed text influences LLM tool usage. * Ingestion points: The on_transcript callback in the Deepgram integration and the input_audio_buffer.append event in the OpenAI Realtime API handle external audio data converted to text. * Boundary markers: The provided code snippets do not include prompt delimiters or specific instructions to ignore malicious commands embedded in user speech. * Capability inventory: The agents are configured with tool-calling capabilities such as get_weather (OpenAI example) and check_order (Vapi webhook example). * Sanitization: The code lacks explicit sanitization or validation of the transcribed text before it is interpolated into the agent's processing logic.

voice-ai-development