The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted audio and video files, which constitutes a vulnerability surface for indirect prompt injection. If an attacker embeds malicious instructions in the audio, the resulting transcription could potentially manipulate the AI agent if it subsequently treats that text as instructions.\n
Ingestion points: Audio and video files are ingested via command-line arguments in scripts/main.py.\n
Boundary markers: The skill does not use delimiters or boundary markers to isolate transcribed text from system-level instructions.\n
Capability inventory: The skill possesses file-write capabilities across its transcription and translation commands.\n
Sanitization: No sanitization or content filtering is performed on the output of the transcription process.\n- [EXTERNAL_DOWNLOADS]: The skill's underlying library, OpenAI Whisper, downloads pre-trained model weights from official sources during its initial run.\n
Source: Downloads originate from OpenAI's official model repositories and are handled by the whisper.load_model function in scripts/main.py.

whisper-transcription