The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection as it ingests untrusted audio/video data and converts it to text which the agent then processes.
Ingestion points: scripts/transcribe.py reads user-provided audio and video files.
Boundary markers: SKILL.md includes instructions for the agent to summarize or answer questions about the transcript rather than outputting it directly, providing a logical boundary.
Capability inventory: The skill script is limited to file reading, transcription, and writing text/subtitle files (.txt, .srt, .vtt). It does not have network-writing or command-execution capabilities beyond the initial setup.
Sanitization: No explicit sanitization of the resulting transcript is performed, but the limited capabilities of the skill minimize the potential impact of an injection.

transcribe-audio