The Agent Skills Directory

[SAFE]: The skill serves as a legitimate architectural guide and implementation template for voice AI systems, using standard Python asynchronous programming practices. All components are designed to run concurrently using asyncio queues, following industry standards for real-time streaming applications.\n- [SAFE]: No hardcoded credentials, sensitive file path access, or unauthorized network operations were identified. External services referenced (OpenAI, Deepgram, Google Cloud, ElevenLabs) are well-known technology providers, and their usage in the templates is purely illustrative for developer integration.\n- [SAFE]: Code examples and documentation are transparent and free of obfuscation techniques, dynamic execution of untrusted code, or suspicious persistence mechanisms.\n- [SAFE]: The interrupt system implementation correctly utilizes threading and asyncio events to manage real-time playback control, which is a required feature for conversational AI and does not present a privilege escalation risk.\n- [INDIRECT_PROMPT_INJECTION]: The skill defines an architecture that is inherently susceptible to indirect prompt injection, as it involves processing untrusted external audio/text data within LLM prompts.\n
Ingestion points: User audio is ingested via WebSockets in examples/complete_voice_engine.py and passed to the LLM agent as text in examples/gemini_agent_example.py.\n
Boundary markers: The templates do not currently implement specific boundary markers or instruction-ignoring delimiters for the user-supplied content.\n
Capability inventory: The system's capabilities are limited to audio processing and LLM interaction; no dangerous primitives like shell execution or file system writing are exposed.\n
Sanitization: Input sanitization is not implemented in the provided templates, which is typical for developer-focused starter code but noteworthy as an attack surface.

voice-ai-engine-development