The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted audio data, which creates a potential attack surface for indirect prompt injection where malicious instructions could be embedded in the audio streams.
Ingestion points: Audio data is ingested from files (e.g., raw/user-question.wav) and passed as input_audio content blocks in the code examples within SKILL.md.
Boundary markers: The skill relies on system-level instructions to guide the model's interpretation and safety, though it lacks strict structural delimiters for audio data.
Capability inventory: The skill possesses the ability to make network requests to OpenAI APIs and perform file system writes for audio and transcript artifacts in the artifacts/ and raw/ directories.
Sanitization: The skill mitigates risks by recommending procedural safeguards, including manual quality reviews, consent verification, and a versioned approval manifest protocol.

openai-audio