mlx-whisper
Pass
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill instructs the user to install the
mlx-whisperPython package and downloads pre-trained model weights from themlx-communityorganization on Hugging Face. These are established resources for running machine learning models on Apple Silicon. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external audio data for transcription, creating a potential surface for indirect prompt injection if the transcribed text is later used to influence agent behavior.
- Ingestion points: Audio files are ingested via the
mlx_whisper.transcribemethod inSKILL.md. - Boundary markers: None are utilized to separate the transcribed content from other instructions.
- Capability inventory: The skill only performs local transcription and prints the results; it does not request capabilities for file system modification, network exfiltration, or shell execution.
- Sanitization: There is no specific sanitization of the audio content or the resulting text transcripts.
Audit Metadata