The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of external audio data and user-supplied instructions, which could be used to deliver malicious commands to the agent.
Ingestion points: Audio files uploaded through the Files API or provided as inline data in Python and JavaScript examples; user-defined prompts for transcription, summarization, and audio analysis.
Boundary markers: The provided code examples lack explicit technical enforcement of boundary markers or delimitation between the user's instructions and the potentially untrusted audio content.
Capability inventory: The skill utilizes the official Google GenAI SDK to perform model inferences and manage files; it also includes script examples that perform file system reads and writes to save generated audio.
Sanitization: The skill includes a 'best practices' guide that provides a sanitize_prompt function to escape double quotes and limit input length, and it encourages the use of safety_settings to filter harmful model outputs.

gemini-audio