parakeet
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill operates entirely locally and does not exhibit any malicious behaviors, obfuscation, or unauthorized network activity.
- [COMMAND_EXECUTION]: The server script (
parakeet_server.py) executes the transcription engine viasubprocess.run. This is implemented securely by using a list of arguments instead of a shell string, ensuring that user-provided file paths are treated as data rather than executable code. - [EXTERNAL_DOWNLOADS]: The documentation describes installing the 'Handy' application via Homebrew and downloading the NVIDIA Parakeet model. These are legitimate external resources required for the skill's functionality and are sourced from official and well-known repositories.
- [INDIRECT_PROMPT_INJECTION]: The skill exposes an attack surface for indirect prompt injection as it processes external audio data.
- Ingestion points: Audio data is ingested via files in
scripts/transcribe.pyand microphone input inscripts/dictate.py. - Boundary markers: The skill does not utilize specific delimiters to separate transcribed content from potential subsequent instructions.
- Capability inventory: The skill is configured with
Bash,Read, andWritetools in its frontmatter. - Sanitization: No sanitization or filtering is applied to the transcribed text before it is returned to the user or agent, which is typical for transcription tools.
Audit Metadata